@azerogluemin/ai-bootstrap 0.2.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +157 -0
- package/LICENSE +21 -0
- package/README.md +146 -0
- package/bin/init.js +10 -0
- package/dist/applier/agents-installer.d.ts +12 -0
- package/dist/applier/agents-installer.js +75 -0
- package/dist/applier/agents-installer.js.map +1 -0
- package/dist/applier/bundle-definitions.d.ts +7 -0
- package/dist/applier/bundle-definitions.js +98 -0
- package/dist/applier/bundle-definitions.js.map +1 -0
- package/dist/applier/index.d.ts +9 -0
- package/dist/applier/index.js +147 -0
- package/dist/applier/index.js.map +1 -0
- package/dist/applier/mcp-catalog.d.ts +29 -0
- package/dist/applier/mcp-catalog.js +755 -0
- package/dist/applier/mcp-catalog.js.map +1 -0
- package/dist/applier/mcp-config.d.ts +29 -0
- package/dist/applier/mcp-config.js +123 -0
- package/dist/applier/mcp-config.js.map +1 -0
- package/dist/applier/profile-writer.d.ts +2 -0
- package/dist/applier/profile-writer.js +60 -0
- package/dist/applier/profile-writer.js.map +1 -0
- package/dist/applier/projects-writer.d.ts +2 -0
- package/dist/applier/projects-writer.js +108 -0
- package/dist/applier/projects-writer.js.map +1 -0
- package/dist/applier/settings-writer.d.ts +2 -0
- package/dist/applier/settings-writer.js +74 -0
- package/dist/applier/settings-writer.js.map +1 -0
- package/dist/applier/skills-installer.d.ts +12 -0
- package/dist/applier/skills-installer.js +90 -0
- package/dist/applier/skills-installer.js.map +1 -0
- package/dist/commands/backup.d.ts +1 -0
- package/dist/commands/backup.js +269 -0
- package/dist/commands/backup.js.map +1 -0
- package/dist/commands/doctor.d.ts +1 -0
- package/dist/commands/doctor.js +185 -0
- package/dist/commands/doctor.js.map +1 -0
- package/dist/commands/mcp.d.ts +1 -0
- package/dist/commands/mcp.js +142 -0
- package/dist/commands/mcp.js.map +1 -0
- package/dist/commands/telemetry.d.ts +1 -0
- package/dist/commands/telemetry.js +85 -0
- package/dist/commands/telemetry.js.map +1 -0
- package/dist/commands/update.d.ts +7 -0
- package/dist/commands/update.js +62 -0
- package/dist/commands/update.js.map +1 -0
- package/dist/index.d.ts +2 -0
- package/dist/index.js +113 -0
- package/dist/index.js.map +1 -0
- package/dist/steps/1-profile.d.ts +2 -0
- package/dist/steps/1-profile.js +81 -0
- package/dist/steps/1-profile.js.map +1 -0
- package/dist/steps/2-projects.d.ts +5 -0
- package/dist/steps/2-projects.js +81 -0
- package/dist/steps/2-projects.js.map +1 -0
- package/dist/steps/3-bundles.d.ts +4 -0
- package/dist/steps/3-bundles.js +95 -0
- package/dist/steps/3-bundles.js.map +1 -0
- package/dist/steps/4-mcps.d.ts +3 -0
- package/dist/steps/4-mcps.js +79 -0
- package/dist/steps/4-mcps.js.map +1 -0
- package/dist/steps/5-memory.d.ts +6 -0
- package/dist/steps/5-memory.js +51 -0
- package/dist/steps/5-memory.js.map +1 -0
- package/dist/steps/6-github.d.ts +5 -0
- package/dist/steps/6-github.js +34 -0
- package/dist/steps/6-github.js.map +1 -0
- package/dist/types.d.ts +57 -0
- package/dist/types.js +3 -0
- package/dist/types.js.map +1 -0
- package/dist/utils/paths.d.ts +10 -0
- package/dist/utils/paths.js +35 -0
- package/dist/utils/paths.js.map +1 -0
- package/dist/utils/permissions.d.ts +15 -0
- package/dist/utils/permissions.js +37 -0
- package/dist/utils/permissions.js.map +1 -0
- package/dist/utils/scanner.d.ts +11 -0
- package/dist/utils/scanner.js +108 -0
- package/dist/utils/scanner.js.map +1 -0
- package/dist/utils/telemetry.d.ts +26 -0
- package/dist/utils/telemetry.js +108 -0
- package/dist/utils/telemetry.js.map +1 -0
- package/dist/wizard.d.ts +2 -0
- package/dist/wizard.js +94 -0
- package/dist/wizard.js.map +1 -0
- package/package.json +77 -0
- package/templates/agents/ai-ml-engineer/AGENT.md +41 -0
- package/templates/agents/analyst-agent/AGENT.md +36 -0
- package/templates/agents/backend-engineer/AGENT.md +85 -0
- package/templates/agents/business-analyst/AGENT.md +47 -0
- package/templates/agents/ceo-agent/AGENT.md +57 -0
- package/templates/agents/cfo-agent/AGENT.md +57 -0
- package/templates/agents/chro-agent/AGENT.md +57 -0
- package/templates/agents/ciso-agent/AGENT.md +57 -0
- package/templates/agents/cmo-agent/AGENT.md +57 -0
- package/templates/agents/code-architect/AGENT.md +69 -0
- package/templates/agents/code-explorer/AGENT.md +67 -0
- package/templates/agents/code-reviewer/AGENT.md +86 -0
- package/templates/agents/competitive-intel/AGENT.md +45 -0
- package/templates/agents/content-calendar-orchestrator/AGENT.md +37 -0
- package/templates/agents/content-marketer/AGENT.md +38 -0
- package/templates/agents/coo-agent/AGENT.md +57 -0
- package/templates/agents/cro-agent/AGENT.md +57 -0
- package/templates/agents/cto-agent/AGENT.md +57 -0
- package/templates/agents/data-engineer/AGENT.md +52 -0
- package/templates/agents/db-migrator/AGENT.md +40 -0
- package/templates/agents/debugger/AGENT.md +64 -0
- package/templates/agents/decision-recorder/AGENT.md +41 -0
- package/templates/agents/designer-agent/AGENT.md +38 -0
- package/templates/agents/devops-engineer/AGENT.md +73 -0
- package/templates/agents/ecommerce-optimizer-agent/AGENT.md +36 -0
- package/templates/agents/education-curriculum/AGENT.md +36 -0
- package/templates/agents/email-marketer/AGENT.md +37 -0
- package/templates/agents/finance-analyst-agent/AGENT.md +36 -0
- package/templates/agents/fintech-compliance/AGENT.md +36 -0
- package/templates/agents/founder-mode-agent/AGENT.md +57 -0
- package/templates/agents/frontend-engineer/AGENT.md +97 -0
- package/templates/agents/gaming-balance-agent/AGENT.md +36 -0
- package/templates/agents/gc-agent/AGENT.md +57 -0
- package/templates/agents/growth-hacker/AGENT.md +37 -0
- package/templates/agents/healthcare-compliance-agent/AGENT.md +36 -0
- package/templates/agents/instagram-orchestrator/AGENT.md +40 -0
- package/templates/agents/launch-orchestrator/AGENT.md +61 -0
- package/templates/agents/legal-researcher-agent/AGENT.md +36 -0
- package/templates/agents/linkedin-orchestrator/AGENT.md +38 -0
- package/templates/agents/market-researcher/AGENT.md +45 -0
- package/templates/agents/performance-profiler/AGENT.md +65 -0
- package/templates/agents/product-manager-agent/AGENT.md +33 -0
- package/templates/agents/product-strategist-agent/AGENT.md +37 -0
- package/templates/agents/publishing-orchestrator/AGENT.md +46 -0
- package/templates/agents/qa-reviewer/AGENT.md +45 -0
- package/templates/agents/real-estate-agent/AGENT.md +36 -0
- package/templates/agents/refactor-planner/AGENT.md +62 -0
- package/templates/agents/research-orchestrator/AGENT.md +37 -0
- package/templates/agents/researcher/AGENT.md +76 -0
- package/templates/agents/security-auditor/AGENT.md +61 -0
- package/templates/agents/seo-aeo/AGENT.md +45 -0
- package/templates/agents/seo-backlinks/AGENT.md +43 -0
- package/templates/agents/seo-brief/AGENT.md +51 -0
- package/templates/agents/seo-competitor/AGENT.md +44 -0
- package/templates/agents/seo-content/AGENT.md +55 -0
- package/templates/agents/seo-drift/AGENT.md +43 -0
- package/templates/agents/seo-ecommerce/AGENT.md +44 -0
- package/templates/agents/seo-google-api/AGENT.md +48 -0
- package/templates/agents/seo-image/AGENT.md +47 -0
- package/templates/agents/seo-international/AGENT.md +43 -0
- package/templates/agents/seo-local/AGENT.md +43 -0
- package/templates/agents/seo-maps/AGENT.md +43 -0
- package/templates/agents/seo-programmatic/AGENT.md +50 -0
- package/templates/agents/seo-schema/AGENT.md +37 -0
- package/templates/agents/seo-semantic/AGENT.md +47 -0
- package/templates/agents/seo-spa/AGENT.md +44 -0
- package/templates/agents/seo-sxo/AGENT.md +43 -0
- package/templates/agents/seo-technical/AGENT.md +56 -0
- package/templates/agents/social-media-manager/AGENT.md +36 -0
- package/templates/agents/sound-designer/AGENT.md +39 -0
- package/templates/agents/stack-tester/AGENT.md +42 -0
- package/templates/agents/storyboard-orchestrator/AGENT.md +36 -0
- package/templates/agents/tiktok-orchestrator/AGENT.md +39 -0
- package/templates/agents/twitter-orchestrator/AGENT.md +38 -0
- package/templates/agents/ux-researcher-agent/AGENT.md +36 -0
- package/templates/agents/video-pipeline/AGENT.md +53 -0
- package/templates/agents/youtube-orchestrator/AGENT.md +40 -0
- package/templates/skills/accessibility-auditor/SKILL.md +129 -0
- package/templates/skills/aeo-specialist/SKILL.md +145 -0
- package/templates/skills/analytics-expert/SKILL.md +128 -0
- package/templates/skills/architect/SKILL.md +258 -0
- package/templates/skills/brand-identity-designer/SKILL.md +156 -0
- package/templates/skills/brand-kit-builder/SKILL.md +182 -0
- package/templates/skills/business-coach/SKILL.md +85 -0
- package/templates/skills/ceo-advisor/SKILL.md +156 -0
- package/templates/skills/cfo-advisor/SKILL.md +146 -0
- package/templates/skills/chaos-engineer/SKILL.md +107 -0
- package/templates/skills/character-designer/SKILL.md +212 -0
- package/templates/skills/character-designer/knowledge/character-anatomy.md +119 -0
- package/templates/skills/character-designer/knowledge/composition.md +94 -0
- package/templates/skills/character-designer/knowledge/consistency-anchors.md +113 -0
- package/templates/skills/character-designer/knowledge/lighting.md +120 -0
- package/templates/skills/chro-advisor/SKILL.md +132 -0
- package/templates/skills/ci-cd-builder/SKILL.md +189 -0
- package/templates/skills/ciso-advisor/SKILL.md +123 -0
- package/templates/skills/cmo-advisor/SKILL.md +192 -0
- package/templates/skills/code-reviewer/SKILL.md +204 -0
- package/templates/skills/color-palette-builder/SKILL.md +143 -0
- package/templates/skills/community-manager/SKILL.md +127 -0
- package/templates/skills/composer/SKILL.md +173 -0
- package/templates/skills/composer/knowledge/arrangement-principles.md +137 -0
- package/templates/skills/composer/knowledge/composition-principles.md +97 -0
- package/templates/skills/composer/knowledge/consistency-anchors.md +149 -0
- package/templates/skills/composer/knowledge/cultural-traditions.md +193 -0
- package/templates/skills/composer/knowledge/dynamics-build-drop.md +148 -0
- package/templates/skills/composer/knowledge/genres/afrobeats.md +73 -0
- package/templates/skills/composer/knowledge/genres/ambient.md +62 -0
- package/templates/skills/composer/knowledge/genres/azerbaijani-mugham.md +90 -0
- package/templates/skills/composer/knowledge/genres/azerbaijani-pop.md +104 -0
- package/templates/skills/composer/knowledge/genres/blues.md +62 -0
- package/templates/skills/composer/knowledge/genres/choral.md +78 -0
- package/templates/skills/composer/knowledge/genres/classical.md +72 -0
- package/templates/skills/composer/knowledge/genres/country.md +64 -0
- package/templates/skills/composer/knowledge/genres/drill.md +71 -0
- package/templates/skills/composer/knowledge/genres/edm.md +75 -0
- package/templates/skills/composer/knowledge/genres/electronic.md +62 -0
- package/templates/skills/composer/knowledge/genres/film-score.md +92 -0
- package/templates/skills/composer/knowledge/genres/folk.md +67 -0
- package/templates/skills/composer/knowledge/genres/hip-hop.md +64 -0
- package/templates/skills/composer/knowledge/genres/house.md +63 -0
- package/templates/skills/composer/knowledge/genres/indie-folk.md +63 -0
- package/templates/skills/composer/knowledge/genres/indie-rock.md +60 -0
- package/templates/skills/composer/knowledge/genres/jazz-fusion.md +60 -0
- package/templates/skills/composer/knowledge/genres/jazz.md +66 -0
- package/templates/skills/composer/knowledge/genres/k-pop.md +81 -0
- package/templates/skills/composer/knowledge/genres/latin.md +88 -0
- package/templates/skills/composer/knowledge/genres/lo-fi.md +64 -0
- package/templates/skills/composer/knowledge/genres/meykhana.md +95 -0
- package/templates/skills/composer/knowledge/genres/orchestral.md +69 -0
- package/templates/skills/composer/knowledge/genres/persian-classical.md +104 -0
- package/templates/skills/composer/knowledge/genres/pop.md +60 -0
- package/templates/skills/composer/knowledge/genres/reggae.md +65 -0
- package/templates/skills/composer/knowledge/genres/rnb.md +62 -0
- package/templates/skills/composer/knowledge/genres/rock.md +61 -0
- package/templates/skills/composer/knowledge/genres/soul.md +63 -0
- package/templates/skills/composer/knowledge/genres/symphonic-mugham.md +92 -0
- package/templates/skills/composer/knowledge/genres/techno.md +63 -0
- package/templates/skills/composer/knowledge/genres/trap.md +61 -0
- package/templates/skills/composer/knowledge/genres/turkish-arabesque.md +87 -0
- package/templates/skills/composer/knowledge/genres/world-music.md +84 -0
- package/templates/skills/composer/knowledge/harmony-basics.md +105 -0
- package/templates/skills/composer/knowledge/instrumentation-guide.md +145 -0
- package/templates/skills/composer/knowledge/instruments/808.md +59 -0
- package/templates/skills/composer/knowledge/instruments/acoustic-guitar.md +46 -0
- package/templates/skills/composer/knowledge/instruments/balaban.md +74 -0
- package/templates/skills/composer/knowledge/instruments/bass.md +50 -0
- package/templates/skills/composer/knowledge/instruments/brass-section.md +60 -0
- package/templates/skills/composer/knowledge/instruments/cello-solo.md +53 -0
- package/templates/skills/composer/knowledge/instruments/choir.md +67 -0
- package/templates/skills/composer/knowledge/instruments/drum-machine.md +49 -0
- package/templates/skills/composer/knowledge/instruments/drums.md +57 -0
- package/templates/skills/composer/knowledge/instruments/electric-guitar.md +57 -0
- package/templates/skills/composer/knowledge/instruments/flute.md +56 -0
- package/templates/skills/composer/knowledge/instruments/kamancha.md +66 -0
- package/templates/skills/composer/knowledge/instruments/nagara.md +68 -0
- package/templates/skills/composer/knowledge/instruments/oud.md +70 -0
- package/templates/skills/composer/knowledge/instruments/piano.md +42 -0
- package/templates/skills/composer/knowledge/instruments/saxophone.md +66 -0
- package/templates/skills/composer/knowledge/instruments/saz.md +75 -0
- package/templates/skills/composer/knowledge/instruments/strings-orchestra.md +68 -0
- package/templates/skills/composer/knowledge/instruments/synth-bass.md +55 -0
- package/templates/skills/composer/knowledge/instruments/synth-lead.md +70 -0
- package/templates/skills/composer/knowledge/instruments/synth-pad.md +61 -0
- package/templates/skills/composer/knowledge/instruments/tar.md +67 -0
- package/templates/skills/composer/knowledge/instruments/trap-hihat.md +50 -0
- package/templates/skills/composer/knowledge/instruments/trumpet.md +63 -0
- package/templates/skills/composer/knowledge/instruments/violin-solo.md +51 -0
- package/templates/skills/composer/knowledge/key-signature-mood.md +128 -0
- package/templates/skills/composer/knowledge/lyrics-by-language.md +79 -0
- package/templates/skills/composer/knowledge/lyrics-craft.md +89 -0
- package/templates/skills/composer/knowledge/melody-craft.md +126 -0
- package/templates/skills/composer/knowledge/moods/aggressive.md +98 -0
- package/templates/skills/composer/knowledge/moods/bittersweet.md +97 -0
- package/templates/skills/composer/knowledge/moods/dreamy.md +98 -0
- package/templates/skills/composer/knowledge/moods/energetic.md +101 -0
- package/templates/skills/composer/knowledge/moods/ethereal.md +99 -0
- package/templates/skills/composer/knowledge/moods/happy.md +93 -0
- package/templates/skills/composer/knowledge/moods/melancholic-hopeful.md +96 -0
- package/templates/skills/composer/knowledge/moods/melancholic.md +99 -0
- package/templates/skills/composer/knowledge/moods/mysterious.md +97 -0
- package/templates/skills/composer/knowledge/moods/nostalgic.md +96 -0
- package/templates/skills/composer/knowledge/moods/peaceful.md +99 -0
- package/templates/skills/composer/knowledge/moods/playful.md +95 -0
- package/templates/skills/composer/knowledge/moods/romantic.md +99 -0
- package/templates/skills/composer/knowledge/moods/tense.md +97 -0
- package/templates/skills/composer/knowledge/moods/triumphant.md +104 -0
- package/templates/skills/composer/knowledge/personas/aygun-kazimova.md +67 -0
- package/templates/skills/composer/knowledge/personas/billie-eilish.md +80 -0
- package/templates/skills/composer/knowledge/personas/bon-iver.md +70 -0
- package/templates/skills/composer/knowledge/personas/daft-punk.md +71 -0
- package/templates/skills/composer/knowledge/personas/hans-zimmer.md +71 -0
- package/templates/skills/composer/knowledge/personas/hildur-gudnadottir.md +72 -0
- package/templates/skills/composer/knowledge/personas/kanye-west.md +75 -0
- package/templates/skills/composer/knowledge/personas/max-martin.md +65 -0
- package/templates/skills/composer/knowledge/personas/neutral-composer.md +66 -0
- package/templates/skills/composer/knowledge/personas/pharrell.md +70 -0
- package/templates/skills/composer/knowledge/personas/rashid-behbudov.md +60 -0
- package/templates/skills/composer/knowledge/personas/sezen-aksu.md +68 -0
- package/templates/skills/composer/knowledge/personas/sufjan-stevens.md +71 -0
- package/templates/skills/composer/knowledge/personas/the-weeknd.md +78 -0
- package/templates/skills/composer/knowledge/production-styles/ambient-textural.md +70 -0
- package/templates/skills/composer/knowledge/production-styles/demo-rough.md +70 -0
- package/templates/skills/composer/knowledge/production-styles/hi-fi.md +69 -0
- package/templates/skills/composer/knowledge/production-styles/live-feel.md +74 -0
- package/templates/skills/composer/knowledge/production-styles/lo-fi.md +55 -0
- package/templates/skills/composer/knowledge/production-styles/maximalist.md +69 -0
- package/templates/skills/composer/knowledge/production-styles/minimal.md +70 -0
- package/templates/skills/composer/knowledge/production-styles/orchestral-cinematic.md +69 -0
- package/templates/skills/composer/knowledge/production-styles/polished-modern.md +56 -0
- package/templates/skills/composer/knowledge/production-styles/vintage-analog.md +67 -0
- package/templates/skills/composer/knowledge/song-structure.md +97 -0
- package/templates/skills/composer/knowledge/tempo-bpm-guide.md +141 -0
- package/templates/skills/composer/knowledge/time-signature.md +124 -0
- package/templates/skills/composer/knowledge/vocal-direction.md +118 -0
- package/templates/skills/composer/knowledge/vocal-styles/duet-male-female.md +71 -0
- package/templates/skills/composer/knowledge/vocal-styles/falsetto-male.md +49 -0
- package/templates/skills/composer/knowledge/vocal-styles/female-alto.md +48 -0
- package/templates/skills/composer/knowledge/vocal-styles/female-mezzo.md +47 -0
- package/templates/skills/composer/knowledge/vocal-styles/female-soprano.md +48 -0
- package/templates/skills/composer/knowledge/vocal-styles/male-baritone.md +44 -0
- package/templates/skills/composer/knowledge/vocal-styles/male-bass.md +45 -0
- package/templates/skills/composer/knowledge/vocal-styles/male-tenor.md +41 -0
- package/templates/skills/composer/knowledge/vocal-styles/mugham-style.md +67 -0
- package/templates/skills/composer/knowledge/vocal-styles/rap-flow.md +75 -0
- package/templates/skills/composer/knowledge/vocal-styles/spoken-word.md +52 -0
- package/templates/skills/composer/knowledge/vocal-styles/whisper.md +55 -0
- package/templates/skills/content-strategist/SKILL.md +106 -0
- package/templates/skills/conversion-optimizer/SKILL.md +98 -0
- package/templates/skills/coo-advisor/SKILL.md +92 -0
- package/templates/skills/copywriter-pro/SKILL.md +151 -0
- package/templates/skills/cro-advisor/SKILL.md +105 -0
- package/templates/skills/cross-platform-strategist/SKILL.md +99 -0
- package/templates/skills/cto-advisor/SKILL.md +166 -0
- package/templates/skills/cultural-translator/SKILL.md +96 -0
- package/templates/skills/decision-maker/SKILL.md +145 -0
- package/templates/skills/director/SKILL.md +236 -0
- package/templates/skills/director/knowledge/camera-movements.md +118 -0
- package/templates/skills/director/knowledge/cameras/cameras.json +234 -0
- package/templates/skills/director/knowledge/cameras/cameras.md +53 -0
- package/templates/skills/director/knowledge/cameras/recommendations.md +75 -0
- package/templates/skills/director/knowledge/cinematography.md +87 -0
- package/templates/skills/director/knowledge/composition.md +94 -0
- package/templates/skills/director/knowledge/lens-selection.md +82 -0
- package/templates/skills/director/knowledge/lenses/lenses.json +292 -0
- package/templates/skills/director/knowledge/lenses/lenses.md +87 -0
- package/templates/skills/director/knowledge/lenses/selection-guide.md +105 -0
- package/templates/skills/director/knowledge/lighting.md +120 -0
- package/templates/skills/director/knowledge/personas/_visual-style-index.md +144 -0
- package/templates/skills/director/knowledge/personas/adam-berg.md +30 -0
- package/templates/skills/director/knowledge/personas/agnes-varda.md +30 -0
- package/templates/skills/director/knowledge/personas/alfonso-cuaron.md +30 -0
- package/templates/skills/director/knowledge/personas/andrea-arnold.md +30 -0
- package/templates/skills/director/knowledge/personas/andrew-stanton.md +33 -0
- package/templates/skills/director/knowledge/personas/apichatpong-weerasethakul.md +30 -0
- package/templates/skills/director/knowledge/personas/ari-aster.md +33 -0
- package/templates/skills/director/knowledge/personas/armando-iannucci.md +30 -0
- package/templates/skills/director/knowledge/personas/asif-kapadia.md +30 -0
- package/templates/skills/director/knowledge/personas/barry-jenkins.md +30 -0
- package/templates/skills/director/knowledge/personas/bob-persichetti.md +33 -0
- package/templates/skills/director/knowledge/personas/bong-joon-ho.md +33 -0
- package/templates/skills/director/knowledge/personas/brad-bird.md +30 -0
- package/templates/skills/director/knowledge/personas/cary-joji-fukunaga.md +30 -0
- package/templates/skills/director/knowledge/personas/chloe-zhao.md +30 -0
- package/templates/skills/director/knowledge/personas/chris-cunningham.md +30 -0
- package/templates/skills/director/knowledge/personas/chris-sanders.md +33 -0
- package/templates/skills/director/knowledge/personas/christopher-mcquarrie.md +30 -0
- package/templates/skills/director/knowledge/personas/coen-brothers.md +30 -0
- package/templates/skills/director/knowledge/personas/dardenne-brothers.md +30 -0
- package/templates/skills/director/knowledge/personas/david-robert-mitchell.md +33 -0
- package/templates/skills/director/knowledge/personas/dean-deblois.md +33 -0
- package/templates/skills/director/knowledge/personas/edgar-wright.md +30 -0
- package/templates/skills/director/knowledge/personas/errol-morris.md +30 -0
- package/templates/skills/director/knowledge/personas/fincher.md +33 -0
- package/templates/skills/director/knowledge/personas/floria-sigismondi.md +30 -0
- package/templates/skills/director/knowledge/personas/frederick-wiseman.md +30 -0
- package/templates/skills/director/knowledge/personas/genndy-tartakovsky.md +33 -0
- package/templates/skills/director/knowledge/personas/george-miller.md +30 -0
- package/templates/skills/director/knowledge/personas/greta-gerwig.md +30 -0
- package/templates/skills/director/knowledge/personas/hayao-miyazaki.md +30 -0
- package/templates/skills/director/knowledge/personas/henry-selick.md +33 -0
- package/templates/skills/director/knowledge/personas/hirokazu-kore-eda.md +30 -0
- package/templates/skills/director/knowledge/personas/hou-hsiao-hsien.md +30 -0
- package/templates/skills/director/knowledge/personas/hype-williams.md +30 -0
- package/templates/skills/director/knowledge/personas/james-cameron.md +33 -0
- package/templates/skills/director/knowledge/personas/jeff-nichols.md +33 -0
- package/templates/skills/director/knowledge/personas/joaquim-dos-santos.md +33 -0
- package/templates/skills/director/knowledge/personas/john-woo.md +30 -0
- package/templates/skills/director/knowledge/personas/jon-favreau.md +33 -0
- package/templates/skills/director/knowledge/personas/jonas-akerlund.md +30 -0
- package/templates/skills/director/knowledge/personas/jonathan-glazer.md +30 -0
- package/templates/skills/director/knowledge/personas/jordan-peele.md +30 -0
- package/templates/skills/director/knowledge/personas/joseph-kosinski.md +33 -0
- package/templates/skills/director/knowledge/personas/kathryn-bigelow.md +30 -0
- package/templates/skills/director/knowledge/personas/kubrick.md +33 -0
- package/templates/skills/director/knowledge/personas/kurosawa.md +33 -0
- package/templates/skills/director/knowledge/personas/lance-acord.md +30 -0
- package/templates/skills/director/knowledge/personas/lanthimos.md +33 -0
- package/templates/skills/director/knowledge/personas/lee-unkrich.md +33 -0
- package/templates/skills/director/knowledge/personas/makoto-shinkai.md +33 -0
- package/templates/skills/director/knowledge/personas/mamoru-hosoda.md +33 -0
- package/templates/skills/director/knowledge/personas/mark-romanek.md +30 -0
- package/templates/skills/director/knowledge/personas/martin-scorsese.md +30 -0
- package/templates/skills/director/knowledge/personas/masaaki-yuasa.md +33 -0
- package/templates/skills/director/knowledge/personas/michael-haneke.md +30 -0
- package/templates/skills/director/knowledge/personas/michel-gondry.md +30 -0
- package/templates/skills/director/knowledge/personas/miguel-sapochnik.md +33 -0
- package/templates/skills/director/knowledge/personas/neutral-director.md +32 -0
- package/templates/skills/director/knowledge/personas/nick-park.md +33 -0
- package/templates/skills/director/knowledge/personas/nicolas-winding-refn.md +33 -0
- package/templates/skills/director/knowledge/personas/nolan.md +33 -0
- package/templates/skills/director/knowledge/personas/nuri-bilge-ceylan.md +33 -0
- package/templates/skills/director/knowledge/personas/park-chan-wook.md +33 -0
- package/templates/skills/director/knowledge/personas/paul-thomas-anderson.md +30 -0
- package/templates/skills/director/knowledge/personas/pawel-pawlikowski.md +33 -0
- package/templates/skills/director/knowledge/personas/pedro-almodovar.md +30 -0
- package/templates/skills/director/knowledge/personas/pendleton-ward.md +33 -0
- package/templates/skills/director/knowledge/personas/pete-docter.md +30 -0
- package/templates/skills/director/knowledge/personas/peter-jackson.md +33 -0
- package/templates/skills/director/knowledge/personas/phil-lord-chris-miller.md +33 -0
- package/templates/skills/director/knowledge/personas/phoebe-waller-bridge.md +30 -0
- package/templates/skills/director/knowledge/personas/pierre-coffin.md +33 -0
- package/templates/skills/director/knowledge/personas/quentin-tarantino.md +30 -0
- package/templates/skills/director/knowledge/personas/rebecca-sugar.md +33 -0
- package/templates/skills/director/knowledge/personas/ridley-scott.md +30 -0
- package/templates/skills/director/knowledge/personas/robert-eggers.md +33 -0
- package/templates/skills/director/knowledge/personas/robert-zemeckis.md +33 -0
- package/templates/skills/director/knowledge/personas/roy-andersson.md +33 -0
- package/templates/skills/director/knowledge/personas/satoshi-kon.md +30 -0
- package/templates/skills/director/knowledge/personas/sean-baker.md +30 -0
- package/templates/skills/director/knowledge/personas/sofia-coppola.md +30 -0
- package/templates/skills/director/knowledge/personas/spike-jonze.md +30 -0
- package/templates/skills/director/knowledge/personas/steven-spielberg.md +30 -0
- package/templates/skills/director/knowledge/personas/taika-waititi.md +30 -0
- package/templates/skills/director/knowledge/personas/terrence-malick.md +30 -0
- package/templates/skills/director/knowledge/personas/tim-burton.md +33 -0
- package/templates/skills/director/knowledge/personas/villeneuve.md +33 -0
- package/templates/skills/director/knowledge/personas/vince-gilligan.md +30 -0
- package/templates/skills/director/knowledge/personas/werner-herzog.md +30 -0
- package/templates/skills/director/knowledge/personas/wes-anderson.md +33 -0
- package/templates/skills/director/knowledge/personas/wong-kar-wai.md +33 -0
- package/templates/skills/director/knowledge/personas/zhang-yimou.md +30 -0
- package/templates/skills/doc-writer/SKILL.md +293 -0
- package/templates/skills/e-commerce-optimizer/SKILL.md +120 -0
- package/templates/skills/elevenlabs/SKILL.md +122 -0
- package/templates/skills/elevenlabs/knowledge/audio-tags.md +52 -0
- package/templates/skills/elevenlabs/knowledge/azerbaijani-voice.md +46 -0
- package/templates/skills/elevenlabs/knowledge/sound-effects.md +131 -0
- package/templates/skills/elevenlabs/knowledge/text-to-speech.md +52 -0
- package/templates/skills/elevenlabs/knowledge/voice-changer.md +39 -0
- package/templates/skills/elevenlabs/knowledge/voice-design.md +100 -0
- package/templates/skills/email-sequence-builder/SKILL.md +174 -0
- package/templates/skills/experiment-designer/SKILL.md +95 -0
- package/templates/skills/finance-analyst/SKILL.md +105 -0
- package/templates/skills/founder-mode/SKILL.md +185 -0
- package/templates/skills/gc-advisor/SKILL.md +146 -0
- package/templates/skills/growth-coach/SKILL.md +95 -0
- package/templates/skills/growth-strategist/SKILL.md +181 -0
- package/templates/skills/healthcare-compliance/SKILL.md +109 -0
- package/templates/skills/image-prompt-engineer/SKILL.md +522 -0
- package/templates/skills/image-prompt-engineer/knowledge/character-anatomy.md +378 -0
- package/templates/skills/image-prompt-engineer/knowledge/cinematography.md +111 -0
- package/templates/skills/image-prompt-engineer/knowledge/color-grading.md +172 -0
- package/templates/skills/image-prompt-engineer/knowledge/composition.md +157 -0
- package/templates/skills/image-prompt-engineer/knowledge/consistency-anchors.md +113 -0
- package/templates/skills/image-prompt-engineer/knowledge/general/aspect-ratios.md +185 -0
- package/templates/skills/image-prompt-engineer/knowledge/general/prompt-budget.md +167 -0
- package/templates/skills/image-prompt-engineer/knowledge/lens-selection.md +82 -0
- package/templates/skills/image-prompt-engineer/knowledge/lighting.md +174 -0
- package/templates/skills/image-prompt-engineer/knowledge/location-design.md +236 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/flux-pro.md +87 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/gpt-image-2.md +120 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/ideogram-v3.md +101 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/imagen-4.md +73 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/midjourney-v7.md +85 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/nano-banana.md +128 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/recraft-v3.md +94 -0
- package/templates/skills/image-prompt-engineer/knowledge/models/sd-3.5.md +90 -0
- package/templates/skills/image-prompt-engineer/knowledge/multi-reference-workflows.md +332 -0
- package/templates/skills/image-prompt-engineer/knowledge/physical-realism.md +604 -0
- package/templates/skills/image-validator/SKILL.md +303 -0
- package/templates/skills/image-validator/knowledge/physical-realism-checks.md +704 -0
- package/templates/skills/image-validator/knowledge/prompt-consistency-checks.md +485 -0
- package/templates/skills/image-validator/knowledge/script-consistency-checks.md +279 -0
- package/templates/skills/image-validator/knowledge/test-learnings.md +1816 -0
- package/templates/skills/image-validator/knowledge/validation-protocol.md +245 -0
- package/templates/skills/inbox-triage/SKILL.md +121 -0
- package/templates/skills/incident-commander/SKILL.md +192 -0
- package/templates/skills/instagram-expert/SKILL.md +159 -0
- package/templates/skills/journal-keeper/SKILL.md +139 -0
- package/templates/skills/knowledge-base-builder/SKILL.md +148 -0
- package/templates/skills/kubernetes-operator/SKILL.md +185 -0
- package/templates/skills/landing-page-builder/SKILL.md +125 -0
- package/templates/skills/learning-keeper/SKILL.md +212 -0
- package/templates/skills/legal-researcher/SKILL.md +112 -0
- package/templates/skills/life-coach/SKILL.md +90 -0
- package/templates/skills/linkedin-expert/SKILL.md +176 -0
- package/templates/skills/location-designer/SKILL.md +227 -0
- package/templates/skills/location-designer/knowledge/color-grading.md +124 -0
- package/templates/skills/location-designer/knowledge/composition.md +94 -0
- package/templates/skills/location-designer/knowledge/lighting.md +120 -0
- package/templates/skills/location-designer/knowledge/location-design.md +236 -0
- package/templates/skills/logo-designer/SKILL.md +113 -0
- package/templates/skills/lyricist/SKILL.md +194 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/ghazal-q/311/231z/311/231l.md +154 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/m/303/274st/311/231zad.md +134 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/m/303/274s/311/231dd/311/231s.md +131 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/m/303/274x/311/231mm/311/231s.md +141 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/q/311/231sid/311/231.md +129 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/rubai.md +104 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/t/311/231rcib/311/231nd.md +151 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/classical-forms/t/311/231rkibb/311/231nd.md +141 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/folk-forms/bayat/304/261.md +198 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/folk-forms/divani.md +171 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/folk-forms/g/311/231rayl/304/261.md +164 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/folk-forms/m/303/274x/311/231mm/311/231s-a/305/237/304/261q.md +195 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/folk-forms/qo/305/237ma.md +171 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/folk-forms/ustadnam/311/231.md +183 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/meter/11-heca-qo/305/237ma.md +128 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/meter/14-heca-divani.md +118 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/meter/7-heca-bayat/304/261-mani.md +133 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/meter/8-heca-g/311/231rayl/304/261.md +128 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/meter/heca-v/311/231zni-overview.md +101 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/meter/modern-s/311/231rb/311/231st.md +166 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/meter//311/231ruz-v/311/231zni-overview.md +145 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/modern-forms/modern-s/311/231rb/311/231st-/305/237eir.md +213 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/modern-forms/pop-mahn/304/261-s/303/266zl/311/231ri.md +240 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/modern-forms/rap-hip-hop-azerbaijani.md +274 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/modern-forms/rock-mahn/304/261-s/303/266zl/311/231ri.md +261 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/anar-na/304/237/304/261lbaz.md +62 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/aq/305/237in-yenisey.md +92 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/a/305/237/304/261q-/311/231l/311/231sg/311/231r.md +84 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/b/311/231xtiyar-vahabzad/311/231.md +85 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/cahangir-cahangirov.md +58 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/f/303/274zuli.md +77 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/host-jamal-ali.md +59 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/host-orxan-zeynall/304/261.md +76 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/host-promete.md +59 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/host-qaraqan.md +62 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/meykhana-style.md +85 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/miri-yusif.md +67 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/modern-pop-lirik.md +85 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/m/303/274/305/237fiq.md +85 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/m/311/231mm/311/231d-araz.md +72 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/neutral-azerbaijani-lyricist.md +102 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/nizami-g/311/231nc/311/231vi.md +66 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/n/311/231simi.md +72 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/qism/311/231t.md +80 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/ramiz-r/303/266v/305/237/311/231n.md +79 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/r/311/231sul-rza.md +86 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/sabir.md +77 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/salam-sarvan.md +79 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/s/311/231m/311/231d-vur/304/237un.md +78 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/tofiq-quliyev.md +58 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/vahid-/311/231ziz.md +68 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/vaqif-bayatl/304/261-od/311/231r.md +60 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/vaqif-s/311/231m/311/231do/304/237lu.md +72 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas/x/311/231tai.md +69 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/personas//305/237/311/231hriyar-del-gerani.md +78 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/techniques/alliterasya.md +221 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/techniques/cinas.md +176 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/techniques/istiar/311/231.md +168 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/techniques/m/303/274raci/311/231t.md +223 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/techniques/t/311/231krir-radif.md +220 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/techniques/t/311/231/305/237beh-metafora.md +161 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/themes/ayr/304/261l/304/261q-h/311/231sr/311/231t.md +227 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/themes/ekzistensial-h/311/231yat.md +213 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/themes/e/305/237q-m/311/231h/311/231bb/311/231t.md +229 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/themes/mistik-irfani.md +197 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/themes/m/303/274asir-h/311/231yat.md +225 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/themes/t/311/231bi/311/231t-f/311/231sill/311/231r.md +200 -0
- package/templates/skills/lyricist/knowledge/azerbaijani/themes/v/311/231t/311/231n-az/311/231rbaycan.md +214 -0
- package/templates/skills/lyricist/knowledge/english/overview.md +104 -0
- package/templates/skills/lyricist/knowledge/english/personas/bob-dylan.md +59 -0
- package/templates/skills/lyricist/knowledge/english/personas/bruce-springsteen.md +62 -0
- package/templates/skills/lyricist/knowledge/english/personas/drake.md +60 -0
- package/templates/skills/lyricist/knowledge/english/personas/frank-ocean.md +62 -0
- package/templates/skills/lyricist/knowledge/english/personas/joni-mitchell.md +60 -0
- package/templates/skills/lyricist/knowledge/english/personas/kendrick-lamar.md +61 -0
- package/templates/skills/lyricist/knowledge/english/personas/leonard-cohen.md +61 -0
- package/templates/skills/lyricist/knowledge/english/personas/neutral-english-lyricist.md +50 -0
- package/templates/skills/lyricist/knowledge/english/personas/phoebe-bridgers.md +61 -0
- package/templates/skills/lyricist/knowledge/english/personas/taylor-swift.md +61 -0
- package/templates/skills/lyricist/knowledge/russian/overview.md +88 -0
- package/templates/skills/lyricist/knowledge/russian/personas/neutral-russian-lyricist.md +47 -0
- package/templates/skills/lyricist/knowledge/russian/personas/okudzhava.md +67 -0
- package/templates/skills/lyricist/knowledge/russian/personas/pushkin.md +65 -0
- package/templates/skills/lyricist/knowledge/russian/personas/viktor-tsoi.md +68 -0
- package/templates/skills/lyricist/knowledge/russian/personas/vysotsky.md +69 -0
- package/templates/skills/lyricist/knowledge/russian/personas/yesenin.md +67 -0
- package/templates/skills/lyricist/knowledge/russian/personas/zemfira.md +69 -0
- package/templates/skills/lyricist/knowledge/turkish/overview.md +89 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/baris-manco.md +66 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/modern-turkish-rap.md +98 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/nazim-hikmet.md +69 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/neutral-turkish-lyricist.md +48 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/orhan-gencebay.md +67 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/sezen-aksu-lyric.md +70 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/tarkan.md +67 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/yunus-emre.md +72 -0
- package/templates/skills/lyricist/knowledge/turkish/personas/zeki-muren.md +65 -0
- package/templates/skills/mcp-server-builder/SKILL.md +226 -0
- package/templates/skills/meeting-notes/SKILL.md +155 -0
- package/templates/skills/migration-architect/SKILL.md +147 -0
- package/templates/skills/multilingual-content/SKILL.md +149 -0
- package/templates/skills/multilingual-copywriter/SKILL.md +206 -0
- package/templates/skills/multilingual-copywriter/knowledge/languages/az.md +179 -0
- package/templates/skills/paid-ads-strategist/SKILL.md +107 -0
- package/templates/skills/process-mapper/SKILL.md +114 -0
- package/templates/skills/product-manager/SKILL.md +98 -0
- package/templates/skills/psychologist/SKILL.md +94 -0
- package/templates/skills/rag-architect/SKILL.md +153 -0
- package/templates/skills/real-estate-analyzer/SKILL.md +131 -0
- package/templates/skills/refactor/SKILL.md +195 -0
- package/templates/skills/screenwriter/SKILL.md +248 -0
- package/templates/skills/screenwriter/knowledge/_visual-style-index.md +76 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/_category.md +172 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/15s-pre-roll.md +137 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/30s-tv-spot.md +171 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/60s-brand-film.md +182 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/6s-bumper.md +110 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/90s-manifesto.md +184 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/comparison-ad.md +207 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/holiday-seasonal.md +223 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/influencer-script.md +191 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/product-demo.md +192 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/retargeting.md +205 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/testimonial.md +198 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/tutorial-ad.md +258 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/ugc-style.md +201 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/formats/unboxing.md +224 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/personas/bernbach.md +101 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/personas/clow.md +107 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/personas/modern-viral.md +137 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/personas/neutral-copywriter.md +146 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/personas/ogilvy.md +97 -0
- package/templates/skills/screenwriter/knowledge/categories/advertising/personas/wieden.md +118 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/_category.md +29 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/andrew-stanton.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/brad-bird.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/dave-callaham.md +30 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/hayao-miyazaki.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/henry-selick.md +30 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/makoto-shinkai.md +33 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/mamoru-hosoda.md +31 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/mamoru-oshii.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/michael-arndt.md +31 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/pete-docter.md +33 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/phil-lord-chris-miller.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/animation/personas/wes-anderson.md +35 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/_category.md +84 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/formats/essay-doc.md +172 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/formats/interview-driven.md +140 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/formats/investigative.md +187 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/formats/personal-doc.md +163 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/formats/verite.md +172 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/asif-kapadia.md +150 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/errol-morris-writer.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/errol-morris.md +143 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/herzog.md +148 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/ken-burns.md +141 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/louis-theroux.md +140 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/neutral-documentarian.md +130 -0
- package/templates/skills/screenwriter/knowledge/categories/documentary/personas/werner-herzog-writer.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/film/_category.md +130 -0
- package/templates/skills/screenwriter/knowledge/categories/film/formats/feature-film.md +241 -0
- package/templates/skills/screenwriter/knowledge/categories/film/formats/short-film-1-5min.md +165 -0
- package/templates/skills/screenwriter/knowledge/categories/film/formats/short-film-5-15min.md +198 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/action.md +218 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/biographical.md +204 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/comedy.md +187 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/crime.md +228 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/drama.md +166 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/experimental.md +210 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/fantasy.md +213 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/horror.md +202 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/magical-realism.md +227 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/mystery.md +224 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/noir.md +237 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/romance.md +215 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/sci-fi.md +200 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/surreal.md +246 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/thriller.md +194 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/war.md +255 -0
- package/templates/skills/screenwriter/knowledge/categories/film/genres/western.md +234 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/alex-garland.md +31 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/bong-joon-ho.md +190 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/coens.md +184 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/coppola-sofia.md +159 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/farhadi.md +208 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/fincher.md +141 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/gerwig.md +148 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/hampton-fancher.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/james-cameron.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/jonze.md +172 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/lanthimos.md +192 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/miyazaki.md +164 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/neutral-screenwriter.md +163 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/nolan.md +103 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/nuri-bilge-ceylan.md +191 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/park-chan-wook.md +174 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/pta.md +180 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/scorsese.md +163 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/tarantino.md +150 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/villeneuve.md +129 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/walsh-boyens.md +32 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/wes-anderson.md +149 -0
- package/templates/skills/screenwriter/knowledge/categories/film/personas/yilmaz-guney.md +202 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/_category.md +71 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/formats/concept-mv.md +146 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/formats/hybrid-mv.md +136 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/formats/lyric-driven.md +169 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/formats/narrative-mv.md +131 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/formats/performance-mv.md +169 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/personas/hype-williams.md +131 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/personas/michel-gondry.md +113 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/personas/neutral-mv-director.md +124 -0
- package/templates/skills/screenwriter/knowledge/categories/music-video/personas/spike-jonze.md +117 -0
- package/templates/skills/screenwriter/knowledge/categories/series/_category.md +156 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/bottle-episode.md +198 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/cold-open.md +195 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/flashback-episode.md +202 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/limited-series-episode.md +196 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/mid-season-episode.md +164 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/mid-season-finale.md +151 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/pilot.md +196 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/season-finale.md +157 -0
- package/templates/skills/screenwriter/knowledge/categories/series/formats/series-finale.md +183 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/david-simon.md +131 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/gilligan.md +124 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/larry-david.md +140 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/lindelof.md +146 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/mike-white.md +172 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/neutral-showrunner.md +150 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/shonda.md +126 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/sorkin.md +169 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/tina-fey.md +151 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/waller-bridge.md +195 -0
- package/templates/skills/screenwriter/knowledge/categories/series/personas/weiner.md +163 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/anthology.md +99 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/documentary-series.md +163 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/kids-animated.md +114 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/prestige-drama.md +76 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/procedural.md +95 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/sitcom-multi-cam.md +100 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/sitcom-single-cam.md +85 -0
- package/templates/skills/screenwriter/knowledge/categories/series/sub-types/soap.md +89 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/_category.md +103 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/before-after.md +134 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/educational-explainer.md +129 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/hook-resolve.md +123 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/mini-vlog.md +118 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/pov-narrative.md +124 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/reel-narrative.md +100 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/tiktok-skit.md +130 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/tiktok-storytime.md +110 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/tiktok-trend.md +126 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/voiceover-listicle.md +147 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/formats/youtube-shorts.md +142 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/personas/atmospheric-cinematic.md +55 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/personas/atmospheric-reel.md +46 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/personas/educational-style.md +111 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/personas/khaby-style.md +109 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/personas/mr-beast-style.md +112 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/personas/neutral-creator.md +123 -0
- package/templates/skills/screenwriter/knowledge/categories/short-form/personas/tiktok-storytime-native.md +118 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/_category.md +81 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/commentary.md +159 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/documentary-style.md +180 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/interview-podcast.md +187 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/personal-story.md +186 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/review.md +176 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/sketch-long.md +164 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/tutorial.md +164 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/video-essay.md +187 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/formats/vlog.md +152 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/personas/casey-neistat.md +128 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/personas/nerdwriter.md +116 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/personas/neutral-youtuber.md +138 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/personas/tom-scott.md +120 -0
- package/templates/skills/screenwriter/knowledge/categories/youtube-long/personas/wendover.md +120 -0
- package/templates/skills/screenwriter/knowledge/core/character-arcs.md +164 -0
- package/templates/skills/screenwriter/knowledge/core/dialogue-craft.md +139 -0
- package/templates/skills/screenwriter/knowledge/core/fountain-format.md +193 -0
- package/templates/skills/screenwriter/knowledge/core/general-storytelling.md +117 -0
- package/templates/skills/screenwriter/knowledge/core/pacing.md +143 -0
- package/templates/skills/screenwriter/knowledge/platforms/cinema-theatrical.md +156 -0
- package/templates/skills/screenwriter/knowledge/platforms/instagram-reel.md +145 -0
- package/templates/skills/screenwriter/knowledge/platforms/streaming-hbo-prestige.md +134 -0
- package/templates/skills/screenwriter/knowledge/platforms/streaming-netflix-style.md +118 -0
- package/templates/skills/screenwriter/knowledge/platforms/tiktok-vertical.md +166 -0
- package/templates/skills/screenwriter/knowledge/platforms/tv-broadcast.md +85 -0
- package/templates/skills/screenwriter/knowledge/platforms/web-banner.md +156 -0
- package/templates/skills/screenwriter/knowledge/platforms/youtube.md +137 -0
- package/templates/skills/security-auditor/SKILL.md +223 -0
- package/templates/skills/seo-optimizer/SKILL.md +130 -0
- package/templates/skills/showrunner/SKILL.md +348 -0
- package/templates/skills/showrunner/knowledge/pipeline.md +164 -0
- package/templates/skills/simplify/SKILL.md +160 -0
- package/templates/skills/social-strategist/SKILL.md +130 -0
- package/templates/skills/storyboard-builder/SKILL.md +367 -0
- package/templates/skills/storyboard-builder/knowledge/composition.md +94 -0
- package/templates/skills/storyboard-builder/knowledge/consistency-anchors.md +113 -0
- package/templates/skills/storyboard-builder/knowledge/continuity-rules.md +152 -0
- package/templates/skills/storyboard-builder/knowledge/temporal-consistency.md +115 -0
- package/templates/skills/suno-prompt-engineer/SKILL.md +330 -0
- package/templates/skills/suno-prompt-engineer/knowledge/aiva.md +93 -0
- package/templates/skills/suno-prompt-engineer/knowledge/mubert.md +104 -0
- package/templates/skills/suno-prompt-engineer/knowledge/prompt-templates-library.md +971 -0
- package/templates/skills/suno-prompt-engineer/knowledge/stable-audio.md +102 -0
- package/templates/skills/suno-prompt-engineer/knowledge/suno-v4.md +125 -0
- package/templates/skills/suno-prompt-engineer/knowledge/suno-v5.md +130 -0
- package/templates/skills/suno-prompt-engineer/knowledge/udio.md +103 -0
- package/templates/skills/suno-prompt-engineer/knowledge/validated-styles-database.md +333 -0
- package/templates/skills/test-writer/SKILL.md +303 -0
- package/templates/skills/tiktok-expert/SKILL.md +147 -0
- package/templates/skills/typography-system-designer/SKILL.md +170 -0
- package/templates/skills/ui-ux-pro-max/SKILL.md +150 -0
- package/templates/skills/ux-researcher/SKILL.md +104 -0
- package/templates/skills/verify/SKILL.md +203 -0
- package/templates/skills/video-prompt-engineer/SKILL.md +331 -0
- package/templates/skills/video-prompt-engineer/knowledge/audio-sync.md +124 -0
- package/templates/skills/video-prompt-engineer/knowledge/camera-movements.md +192 -0
- package/templates/skills/video-prompt-engineer/knowledge/continuity-rules.md +152 -0
- package/templates/skills/video-prompt-engineer/knowledge/duration-pacing.md +130 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/hailuo.md +78 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/kling-2.md +280 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/luma.md +97 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/pika-2.md +76 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/runway-gen-4.md +102 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/seedance-2.md +80 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/sora.md +116 -0
- package/templates/skills/video-prompt-engineer/knowledge/models/veo-3.md +95 -0
- package/templates/skills/video-prompt-engineer/knowledge/motion-vs-static.md +229 -0
- package/templates/skills/video-prompt-engineer/knowledge/physical-realism.md +369 -0
- package/templates/skills/video-prompt-engineer/knowledge/physics-rules.md +147 -0
- package/templates/skills/video-prompt-engineer/knowledge/subject-motion.md +220 -0
- package/templates/skills/video-prompt-engineer/knowledge/temporal-consistency.md +215 -0
- package/templates/skills/youtube-expert/SKILL.md +196 -0
- package/templates/skills/youtube-thumbnail-designer/SKILL.md +631 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/color-strategies.md +180 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/ctr-formulas.md +215 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/host-reference-generation.md +193 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/icon-and-graphic-elements.md +201 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/mobile-legibility.md +189 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/professional-techniques.md +507 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/safe-zones.md +172 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/topic-enrichment-research.md +241 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/typography.md +194 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/variant-differentiation.md +262 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/world-class-design-baseline.md +333 -0
- package/templates/skills/youtube-thumbnail-designer/knowledge/youtube-analysis.md +289 -0
|
@@ -0,0 +1,1816 @@
|
|
|
1
|
+
# Test Learnings — Failure Pattern Log
|
|
2
|
+
|
|
3
|
+
**Məqsəd:** Hər image validation iterasiyasında aşkar olunan failure pattern-lərini, root cause-larını, və işləyən fix-ləri burada saxla. Sistem **iterasiya ilə öyrənir** — eyni səhv ikinci dəfə təkrar olunmasın deyə yeni knowledge bu fayla əlavə olunur.
|
|
4
|
+
|
|
5
|
+
**İstifadə qaydası:**
|
|
6
|
+
- Validator hər failed test üçün **yeni entry əlavə edir** (chronological)
|
|
7
|
+
- Image-prompt-engineer fix mode-a girəndə bu faylı oxuyur — keçmiş eyni pattern varsa, sınanmış fix-i tətbiq edir
|
|
8
|
+
- Pattern 3+ test-də təkrar olunarsa, master `physical-realism.md` faylına **rule olaraq promote** olunur
|
|
9
|
+
|
|
10
|
+
---
|
|
11
|
+
|
|
12
|
+
## Entry şablonu
|
|
13
|
+
|
|
14
|
+
```markdown
|
|
15
|
+
## YYYY-MM-DD — Test ID (məs. P-03 iter 1→2 / Project cell-3 v1→v2)
|
|
16
|
+
|
|
17
|
+
### Failure (orijinal)
|
|
18
|
+
**Prompt parça:** "..."
|
|
19
|
+
**Beklənilən nəticə:** ...
|
|
20
|
+
**Faktiki nəticə:** ...
|
|
21
|
+
**❌ Pozulma kateqoriyası:** [physical-realism bölmə X / prompt-consistency F-N / identity match]
|
|
22
|
+
|
|
23
|
+
### Root cause analizi
|
|
24
|
+
- AI bias / failure mode: ...
|
|
25
|
+
- Niyə model belə render edib: ...
|
|
26
|
+
- Hansı statistic association qırılıb: ...
|
|
27
|
+
|
|
28
|
+
### Fix tətbiqi
|
|
29
|
+
**Yeni prompt strategiyası:**
|
|
30
|
+
- ...
|
|
31
|
+
**İşlədi / işləmədi:** ...
|
|
32
|
+
|
|
33
|
+
### Carry-forward rule (gələcək prompt-lar üçün)
|
|
34
|
+
- ...
|
|
35
|
+
|
|
36
|
+
### Knowledge update tövsiyəsi
|
|
37
|
+
- [ ] `physical-realism.md` bölmə X-ə əlavə et: ...
|
|
38
|
+
- [ ] `prompt-consistency-checks.md` yeni failure pattern F-N əlavə et: ...
|
|
39
|
+
- [ ] Master master sync (sync rule)
|
|
40
|
+
```
|
|
41
|
+
|
|
42
|
+
---
|
|
43
|
+
|
|
44
|
+
## 2026-05-15 — P-01 (Multi-Light Indoor Conflict)
|
|
45
|
+
|
|
46
|
+
### Failure (orijinal)
|
|
47
|
+
**Prompt parça:** "At midday (12:00 PM, bright sunlight outside). Large window on the camera-left wall lets in strong direct sunlight."
|
|
48
|
+
**Beklənilən nəticə:** Qısa, sharp, neutral white-yellow shadows (sun directly overhead).
|
|
49
|
+
**Faktiki nəticə:** Uzun, warm, golden hour shadows (saatdan asılı olmayaraq).
|
|
50
|
+
**❌ Pozulma kateqoriyası:** physical-realism bölmə 3 (Time-of-day consistency) + prompt-consistency F8 (time-of-day drift)
|
|
51
|
+
|
|
52
|
+
### Root cause analizi
|
|
53
|
+
- AI bias: **"Golden hour default"** — modelin training data-sında landscape/interior şəkillərinin çoxu golden hour-da çəkilib. Model "atmospheric" və "cinematic" lookları golden hour kimi rendered edir.
|
|
54
|
+
- Statistical association: "indoor + window sunlight" → 80% golden hour samples, 15% midday, 5% other
|
|
55
|
+
- "Midday" sözü təkbaşına override etmir — model onu "metaphorical" qəbul edir
|
|
56
|
+
|
|
57
|
+
### Fix tətbiqi
|
|
58
|
+
**Yeni prompt strategiyası:**
|
|
59
|
+
- Explicit numeric time: "Midday at 12:00 noon"
|
|
60
|
+
- Explicit sun position: "sun directly OVERHEAD"
|
|
61
|
+
- Explicit shadow geometry: "Shadows are SHORT (max 30% of object height)"
|
|
62
|
+
- Explicit color temp: "NEUTRAL white-yellow daylight (5600-6000K)"
|
|
63
|
+
- Anti-default override: "NOT warm orange, NOT golden hour"
|
|
64
|
+
- Verify clause: "Verify shadow length against object"
|
|
65
|
+
|
|
66
|
+
**İşlədi:** Daha qısa shadow + neutral colour əldə olundu (lakin tam neutral deyildi — biraz warm tint qaldı, model bias hələ də güclü).
|
|
67
|
+
|
|
68
|
+
### Carry-forward rule
|
|
69
|
+
**For midday/noon scenes:**
|
|
70
|
+
> Explicit numeric time + sun position + shadow geometry + color temp + anti-default override + verify clause. Tək "midday" sözü yetərli deyil.
|
|
71
|
+
|
|
72
|
+
### Knowledge update tövsiyəsi
|
|
73
|
+
- [x] `physical-realism.md` bölmə 3 (Time of day) — "Midday lighting override template" əlavə olundu
|
|
74
|
+
- [x] `prompt-consistency-checks.md` F8 (Time-of-day drift) — template güncəlləndi
|
|
75
|
+
|
|
76
|
+
---
|
|
77
|
+
|
|
78
|
+
## 2026-05-15 — P-02 (Complex Clothesline Sag)
|
|
79
|
+
|
|
80
|
+
### Failure (orijinal)
|
|
81
|
+
**Prompt parça:** "The clothesline SAGS realistically under weight — it dips LOWER under the heavy bath towel and jeans, and rises HIGHER between the lighter items."
|
|
82
|
+
**Beklənilən nəticə:** Görünən catenary curve (V-şəkilli sag), heavy item-lər altında belin enməsi.
|
|
83
|
+
**Faktiki nəticə:** İp demək olar düz, kiçik sag (subtle), heavy/light differential görünmür.
|
|
84
|
+
**❌ Pozulma kateqoriyası:** physical-realism bölmə 1.3 (Yüklə uyğun sag/bend)
|
|
85
|
+
|
|
86
|
+
### Root cause analizi
|
|
87
|
+
- AI bias: **"Tense rope default"** — modelin training data-sında çəkilmiş ip-lər (yarış xətti, ticarət ip-ləri, photo studio backdrops) əksəriyyət tense və düz çəkilir. Drying line samples az.
|
|
88
|
+
- "Sags realistically" subjektiv termin — model "subtle sag" interpretasiya edir (drama-sız)
|
|
89
|
+
- Heavy/light differential information modelə differential rendering kimi keçmir — model bütün span-ı eyni nominal sag-la rendered edir
|
|
90
|
+
|
|
91
|
+
### Fix tətbiqi
|
|
92
|
+
**Yeni prompt strategiyası:**
|
|
93
|
+
- Aggressive sag language: "VISIBLE DRAMATIC SAG, V-shape curve"
|
|
94
|
+
- Numeric differential: "Lowest point under bath towel sags 15-20cm below the endpoints"
|
|
95
|
+
- Comparative differential: "Section between handkerchief and T-shirt sags ONLY 3-5cm (much less)"
|
|
96
|
+
- Anti-default: "NOT a straight tight rope, NOT subtle sag — dramatic V-shape required"
|
|
97
|
+
|
|
98
|
+
**İşlədi:** Daha visible sag əldə olundu, lakin V-shape hələ də soft idi. **3+ iterasiya** lazımdır tam dramatic effect üçün.
|
|
99
|
+
|
|
100
|
+
### Carry-forward rule
|
|
101
|
+
**For any "loaded rope/wire/cable" prompt:**
|
|
102
|
+
> Numeric sag depth + comparative depth across sections + anti-default ("NOT straight, NOT tight, NOT subtle") + V-shape geometry language. Without numeric and comparative, model defaults to nearly-straight.
|
|
103
|
+
|
|
104
|
+
### Knowledge update tövsiyəsi
|
|
105
|
+
- [x] `physical-realism.md` bölmə 1.3 (Yüklə uyğun sag/bend) — "Catenary sag emphasis template" əlavə olundu
|
|
106
|
+
- [ ] `prompt-consistency-checks.md` yeni F9 əlavə et: "Subjective physics quantifier failure" — modelin "realistic", "natural", "dramatic" sözlərini subtle interpretasiya etməsi
|
|
107
|
+
|
|
108
|
+
---
|
|
109
|
+
|
|
110
|
+
## 2026-05-15 — P-03 (Spiral Staircase — iteration 1 → 2)
|
|
111
|
+
|
|
112
|
+
### Failure (orijinal, iteration 1)
|
|
113
|
+
**Prompt parça:** "The INNER SIDE of the spiral is the central post (handhold via attached helical iron handrail spiraling up)."
|
|
114
|
+
**Beklənilən nəticə:** Solid central post + inner helical handrail attached by brackets, hand-height.
|
|
115
|
+
**Faktiki nəticə:** Central post ✅, BUT inner handrail tamamən rendered olunmayıb (yox).
|
|
116
|
+
**❌ Pozulma kateqoriyası:** physical-realism bölmə 7.3 (Pilləkən struktur) + prompt-consistency yeni F9 (dual element rendering bias)
|
|
117
|
+
|
|
118
|
+
### Root cause analizi
|
|
119
|
+
- **AI bias: Dual-element simplification** — model "spiral staircase" prompt-da **bir banister sistemi** render edir (statistic norm). İki banister sistemi (inner + outer) **modelin training data-sında nadir** kombinasiyadır.
|
|
120
|
+
- "handhold via attached helical iron handrail" parantezdə yazılıb → model parantezdəki bilgini **opsional** kimi qəbul edib
|
|
121
|
+
- "Inner side IS the central post" konfliktli (post tək struktur kimi rendered olunur, "second rail attached to it" instruction-u modelə subsequent thought kimi gəlir, və attention drop edilir)
|
|
122
|
+
|
|
123
|
+
### Fix tətbiqi (iteration 1 → 2)
|
|
124
|
+
**Yeni prompt strategiyası:**
|
|
125
|
+
- Numbered subsystem headers: **"TWO SEPARATE BANISTER SYSTEMS — both MUST be rendered"** + "3A. OUTER" + "3B. INNER"
|
|
126
|
+
- Explicit attachment language: "attached BY VISIBLE METAL BRACKETS (small L-shaped arms protruding from the column every ~1 meter)"
|
|
127
|
+
- Safety/fail language: "Without this, the prompt fails: ascending a tall spiral with NO inner handhold is physically dangerous and architecturally incorrect"
|
|
128
|
+
- Verification checklist at end: "- [ ] INNER iron handrail attached to central column by visible brackets (count brackets)"
|
|
129
|
+
- "Verify by checking..." clause
|
|
130
|
+
|
|
131
|
+
**İşlədi:** ✅ İnner handrail bracket-lərlə bərkidilmiş halda görünür. Outer banister də bütün turn-lərdə clear. Both systems rendered.
|
|
132
|
+
|
|
133
|
+
### Carry-forward rule
|
|
134
|
+
**For any "dual or multi-element" prompt (multiple banisters, multiple lights, multiple supports, multiple rails, etc.):**
|
|
135
|
+
1. Use numbered subsystem headers (3A, 3B, 3C, etc.) — force model to allocate attention to each element separately
|
|
136
|
+
2. Each subsystem requires its own attachment/connection description
|
|
137
|
+
3. "TWO/THREE SEPARATE [element]s — ALL MUST be rendered" as header clause
|
|
138
|
+
4. Verification checklist at end listing each element
|
|
139
|
+
5. Avoid parenthetical mention of secondary elements — model treats parens as optional
|
|
140
|
+
|
|
141
|
+
**Generalized to F9 failure pattern (new):**
|
|
142
|
+
> **F9 — Dual/Multi-element rendering bias:** Model defaults to rendering ONE instance of similar elements (one banister, one light, one support) even when prompt mentions multiple. Fix: numbered subsystems + explicit "ALL must be rendered" + per-element verification clause.
|
|
143
|
+
|
|
144
|
+
### Knowledge update tövsiyəsi
|
|
145
|
+
- [x] `physical-realism.md` bölmə 7.3 (Pilləkən struktur) — "Dual-banister rule" yeni alt-bölmə kimi əlavə olunmalı
|
|
146
|
+
- [x] `prompt-consistency-checks.md` yeni F9 əlavə olunmalı (yuxarıda yazılıb)
|
|
147
|
+
- [x] Bu fayl (test-learnings.md) yaradıldı və P-01, P-02, P-03 entries yazıldı
|
|
148
|
+
|
|
149
|
+
---
|
|
150
|
+
|
|
151
|
+
## 2026-05-15 — P-04 (Two People Shaking Hands) — SUCCESS (1-shot, no iteration)
|
|
152
|
+
|
|
153
|
+
### Test context
|
|
154
|
+
**Test edir:** Anatomy (hands, finger count) + Material physics (clothing drape, watch, table)
|
|
155
|
+
**Risk pattern:** AI hand generation tipik failure mode-u (fused fingers, 4/6 fingers, morphed). Top AI failure point.
|
|
156
|
+
|
|
157
|
+
### Pattern application (proaktiv tətbiq)
|
|
158
|
+
Bu test, P-01/P-02/P-03 öyrənmələrini **proaktiv** tətbiq etdi (failure-dan sonra deyil, failure-dan əvvəl):
|
|
159
|
+
|
|
160
|
+
1. **F6 (number ignored) → tətbiq olundu:**
|
|
161
|
+
- "EXACTLY 5 fingers", "10 fingers total"
|
|
162
|
+
- "Verify by counting before finalizing"
|
|
163
|
+
|
|
164
|
+
2. **F9 (dual-element bias) → tətbiq olundu:**
|
|
165
|
+
- "1A. HAND A" + "1B. HAND B" numbered subsystems
|
|
166
|
+
- "TWO SEPARATE HANDS must both be" header clause
|
|
167
|
+
- Per-element verification: "Hand A: count fingers. Hand B: count fingers."
|
|
168
|
+
|
|
169
|
+
3. **Anti-default override → tətbiq olundu:**
|
|
170
|
+
- "NOT 4 fingers per hand, NOT 6 fingers per hand"
|
|
171
|
+
- "NOT fused, NOT melted, NOT morphed"
|
|
172
|
+
- "NO fingers passing through each other"
|
|
173
|
+
|
|
174
|
+
4. **Parenthetical drop fix → tətbiq olundu:**
|
|
175
|
+
- Hər vacib element ayrı numbered subitem (sub-bullets parantezdə yox)
|
|
176
|
+
|
|
177
|
+
5. **Verification checklist at end → tətbiq olundu:**
|
|
178
|
+
- 6-bənd verify checklist promptun sonunda
|
|
179
|
+
- "If ANY of the above fails — regenerate"
|
|
180
|
+
|
|
181
|
+
### Nəticə
|
|
182
|
+
**30 checkpoint-dan 30-u ✅** — heç bir ❌ və ya ⚠️. **First-shot success**.
|
|
183
|
+
|
|
184
|
+
Tipik AI hand bug-larından (fused, 4/6 fingers, morphed) heç biri baş vermədi.
|
|
185
|
+
|
|
186
|
+
### Confirmation: patterns work proactively
|
|
187
|
+
P-04 göstərir ki, **failure-dan əvvəl** test-learnings.md tətbiqi **iteration sayını azaldır** (1 shot vs 2-3 shot). Bu, sistemin əsl dəyəridir — yalnız retroactive fix-lər yox, **proaktiv prevention**.
|
|
188
|
+
|
|
189
|
+
### Knowledge update tövsiyəsi
|
|
190
|
+
- [x] Bu success case test-learnings.md-də qeydə alındı
|
|
191
|
+
- [ ] Sonrakı tələb: hər prompt-yazma session-ında image-prompt-engineer ÖZÜ test-learnings.md-i Read etsin (yalnız Fix mode-da yox, Initial mode-da da) — failure prevention universal
|
|
192
|
+
- Tövsiyə CLAUDE.md-də qeyd olunsun: image-prompt-engineer-in **Initial mode**-da test-learnings.md proactive consultation tələb olunsun
|
|
193
|
+
|
|
194
|
+
---
|
|
195
|
+
|
|
196
|
+
## 2026-05-15 — P-05 (Window Reflection) — iter 1 FAILURE
|
|
197
|
+
|
|
198
|
+
### Failure (orijinal)
|
|
199
|
+
**Prompt parça:** "2A. TRANSPARENCY LAYER... 2B. REFLECTION LAYER... 2C. LAYER INTERACTION..."
|
|
200
|
+
**Beklənilən nəticə:** Glass surface-də **iki layer simultaneously** — woman's mirrored silhouette + warm lamp glow IN REFLECTION + city visible THROUGH glass. Fresnel layered.
|
|
201
|
+
**Faktiki nəticə:** Yalnız transparency layer rendered olub. Woman's reflection yoxdur. Lamp rendered olub, lakin **city scene-ə inteqrasiya** olunub (sanki uzaq binanın interyeri), glass səthində layered reflection kimi DEYIL.
|
|
202
|
+
**❌ Pozulma kateqoriyası:** physical-realism A6.2 (glass material) + A8.3 (reflection physics) + prompt-consistency F9 (dual-element bias)
|
|
203
|
+
|
|
204
|
+
### Root cause analizi
|
|
205
|
+
- **AI bias: Single-layer surface rendering** — Model şüşə səthini ya **tamamilə transparent** ya da **tamamilə mirror** kimi render edir. **Hybrid Fresnel layering** (eyni səthdə iki layer eyni anda) zəif statistical association.
|
|
206
|
+
- **Spatial integration failure** — Promptdə "in reflection" yazıldı, lakin model lamp-ı transparent city scene-ə spatial olaraq inteqrasiya etdi (uzaq bina interyeri kimi)
|
|
207
|
+
- F9 generalization: Dual-element bias **physical objects**-də işlədikcə işlədi (P-03 banister, P-04 hands), lakin **dual-LAYER-on-same-surface** üçün daha güclü override lazımdır
|
|
208
|
+
|
|
209
|
+
### Fix tətbiqi (iteration 2 planlanan)
|
|
210
|
+
**Yeni strategiya:**
|
|
211
|
+
1. **Layer explicit positioning:** "Reflection layer is at the GLASS PLANE depth, NOT at city depth. Imagine a flat 2D layer on the glass surface itself."
|
|
212
|
+
2. **Anti-default ultra-aggressive:** "NOT a streetlight in the city, NOT a window in distant buildings — the lamp glow is REFLECTED on the near surface of the glass, floating visually on the glass plane."
|
|
213
|
+
3. **Compositional clue:** "Woman's reflection appears at the SAME 2D position in the frame where her body is — overlaid as a translucent ghost on top of her back."
|
|
214
|
+
4. **Numerical position:** "Reflection occupies approximately the same screen-area as the foreground room/woman, mirrored left-right."
|
|
215
|
+
5. **Reference language:** "Like a window in a movie scene where you see both the city AND the character's face reflected in the glass at the same time."
|
|
216
|
+
|
|
217
|
+
### Carry-forward rule (gələcək glass/mirror prompts üçün)
|
|
218
|
+
**Glass/mirror surface dual-layer rule:**
|
|
219
|
+
> Glass və ya mirror surface-də iki layer (transparency + reflection) eyni anda rendered olunmalıdırsa:
|
|
220
|
+
> 1. **Spatial depth explicit qoy** — reflection AT THE GLASS PLANE, not at scene depth
|
|
221
|
+
> 2. **Anti-default ultra-aggressive** — "NOT a streetlight, NOT a distant window, NOT an outdoor lamp" — modelin "city integration" bias-ını qır
|
|
222
|
+
> 3. **Position match instruction** — reflected object position kadrda harada olmalıdır (məs. "reflection occupies same screen area as foreground")
|
|
223
|
+
> 4. **Reference movie/cinematography language** — "like a film scene where..."
|
|
224
|
+
> 5. **Compositional ghost language** — "translucent ghost overlay" reflection üçün daha yaxşı işləyir model üçün
|
|
225
|
+
|
|
226
|
+
### Yeni pattern: F10 — Surface dual-layer bias
|
|
227
|
+
**Yeni failure pattern (Fresnel-spesifik):**
|
|
228
|
+
> **F10 — Glass/mirror surface dual-layer bias:** Model dual layer on same surface rendering-də yalnız bir layer-ə fokuslanır (adətən transparency, reflection drop). Fix: spatial depth explicit + anti-default ultra + position match + cinema reference language.
|
|
229
|
+
|
|
230
|
+
### Knowledge update tövsiyəsi
|
|
231
|
+
- [x] Test-learnings.md-ə bu entry əlavə olundu
|
|
232
|
+
- [x] prompt-consistency-checks.md-ə F10 əlavə olundu
|
|
233
|
+
- [ ] physical-realism.md bölmə 6.1/6.2 (material — şüşə) "Fresnel dual-layer rendering" alt-bölmə əlavə olunsun
|
|
234
|
+
- [ ] physical-realism-checks.md (sinxron) A6.2 yenilənsin
|
|
235
|
+
|
|
236
|
+
### İteration 2 nəticə — SUCCESS
|
|
237
|
+
**P-05 v2 generation result:** 28/28 ✅ — bütün ❌-lar ✅-yə çevrildi.
|
|
238
|
+
|
|
239
|
+
**Hansı texnikalar həlledici işlədi:**
|
|
240
|
+
1. **"Lost in Translation" cinema reference** — model film aesthetic-i tanıdı və layered reflection sxemini doğru render etdi
|
|
241
|
+
2. **"Translucent ghost" + "double-exposure" language** — reflection-ın overlay təbiətini modelə dəqiq başa saldı
|
|
242
|
+
3. **"AT THE GLASS PLANE depth, NOT at city depth"** — spatial separation kritik idi
|
|
243
|
+
4. **3 anti-default NOT-clause** ("NOT streetlight, NOT distant window, NOT outdoor lamp") — modelin city-integration bias-ı qırıldı
|
|
244
|
+
5. **Failure-conditional verification** ("If reflected lamp blends INTO city → FAILURE → regenerate") — modelə açıq xəbərdarlıq
|
|
245
|
+
|
|
246
|
+
**F10 pattern confirmed:** Glass/mirror surface dual-layer rendering — bu kombinasiya işləyir, gələcək glass/mirror prompts üçün template hazırdır.
|
|
247
|
+
|
|
248
|
+
---
|
|
249
|
+
|
|
250
|
+
## 2026-05-15 — P-06 (Rainy Night Street) — SUCCESS (1-shot, proactive prevention)
|
|
251
|
+
|
|
252
|
+
### Test context
|
|
253
|
+
**Test edir:** Weather physics (rain direction, droplet behavior) + dual-element (umbrella zones) + Fresnel surface (wet asphalt)
|
|
254
|
+
**Risk pattern:** F9 (dual-element — umbrella protected zone), F10 (Fresnel — wet asphalt), F8 (time-of-day rigor — night sodium light)
|
|
255
|
+
|
|
256
|
+
### Pattern application (proactive — fərqli pattern-lər birlikdə)
|
|
257
|
+
Bu test **üç proaktiv pattern** eyni promptda tətbiq etdi:
|
|
258
|
+
|
|
259
|
+
1. **F9 (dual-element) — umbrella zones:**
|
|
260
|
+
- "5A. UNDER the umbrella (protected dry zone)" + "5B. OUTSIDE the umbrella's silhouette (rain zone)"
|
|
261
|
+
- Sharp boundary instruction
|
|
262
|
+
- Per-zone description (dry vs wet)
|
|
263
|
+
|
|
264
|
+
2. **F10 (surface dual-layer) — wet asphalt:**
|
|
265
|
+
- "Fresnel reflection scene on the asphalt" header
|
|
266
|
+
- Roger Deakins "Blade Runner 2049" cinema reference
|
|
267
|
+
- Spatial logic: surface + reflection layered
|
|
268
|
+
|
|
269
|
+
3. **F8 (time-of-day rigor):**
|
|
270
|
+
- Numeric color temp "sodium 2200K orange-yellow"
|
|
271
|
+
- Anti-default "NOT cool blue moonlight"
|
|
272
|
+
|
|
273
|
+
4. **3 numbered LAYER blocks** (atmospheric, ground, light source) — separate spatial logic
|
|
274
|
+
|
|
275
|
+
### Nəticə
|
|
276
|
+
**34 checkpoint-dan 34-u keçdi (32 ✅, 2 ⚠️ minor):**
|
|
277
|
+
- ⚠️ Rain angle daha vertical idi (10° aydın deyil) — minor, scene logic-də işləyir
|
|
278
|
+
- ⚠️ Pant wet verify edilə bilmədi (dark scene) — vizual məhdudiyyət, lakin scene logic-də inconsistent deyil
|
|
279
|
+
|
|
280
|
+
**Heç bir ❌ yoxdur. 1-shot success.**
|
|
281
|
+
|
|
282
|
+
### Confirmation: multi-pattern proactive works
|
|
283
|
+
P-06, **birdən artıq pattern eyni promptda** proactive tətbiqinin işlədiyini sübut etdi. Pattern-lər conflict yaratmır — onlar **complementary** işləyir:
|
|
284
|
+
- F9 element-ləri ayırır
|
|
285
|
+
- F10 səth fizikasını dəqiqləşdirir
|
|
286
|
+
- F8 vaxt baseline-ı qurur
|
|
287
|
+
|
|
288
|
+
### Bonus: Cinema reference güclü vasitə kimi
|
|
289
|
+
İndi 3 test-də (P-05 Sofia Coppola, P-06 Roger Deakins) cinema reference **kritik fix tool** kimi işlədi. Pattern aşkar:
|
|
290
|
+
|
|
291
|
+
**Cinema reference pattern (universal):**
|
|
292
|
+
> Hər kompleks vizual situation üçün uyğun film/director reference əlavə et. Model film aesthetics-i statistical olaraq güclü association saxlayır (training data-da çox film imagery var). Bu reference müəyyən mood + lighting + composition + physics-i bir paket kimi modelə təqdim edir.
|
|
293
|
+
|
|
294
|
+
### Knowledge update tövsiyəsi
|
|
295
|
+
- [x] Test-learnings.md-ə bu entry əlavə olundu
|
|
296
|
+
- [ ] Yeni "Cinema reference as fix tool" pattern test-learnings.md statistical insights-də qeyd olunsun
|
|
297
|
+
- [ ] Possible new pattern "F11 — Cinema reference power-up" (3+ test-də işlədi, promote candidate)
|
|
298
|
+
|
|
299
|
+
---
|
|
300
|
+
|
|
301
|
+
## Statistical insights (3 test sonra)
|
|
302
|
+
|
|
303
|
+
### Pattern: AI subjective quantifier weakness
|
|
304
|
+
P-01 ("midday") + P-02 ("realistically sags") + P-03 ("attached handrail") — hər üçündə **subjective term** model tərəfindən zəif execute olundu. Model **numeric/aggressive/verify-clause** kombinasiyasına daha yaxşı cavab verir.
|
|
305
|
+
|
|
306
|
+
**Rule (universal):**
|
|
307
|
+
> Hər vacib instruction üçün:
|
|
308
|
+
> 1. **Numeric specificity** (cm, percent, count)
|
|
309
|
+
> 2. **Anti-default override** ("NOT X, NOT Y")
|
|
310
|
+
> 3. **Verify clause** ("Verify by checking Z")
|
|
311
|
+
|
|
312
|
+
### Pattern: AI parenthetical drop
|
|
313
|
+
P-03 v1-də parantezdə yazılan "handhold via attached helical iron handrail" tamamilə rendered olunmadı. Model parantezi "opsional clarification" kimi qəbul edir.
|
|
314
|
+
|
|
315
|
+
**Rule:**
|
|
316
|
+
> Heç bir vacib instruction-u parantezdə yazma. Hər vacib element ya numbered subsystem-də, ya da müstəqil cümlədə olsun.
|
|
317
|
+
|
|
318
|
+
### Pattern: Multiple iteration needed for "physics drama"
|
|
319
|
+
P-02 (dramatic sag) — 1 fix iteration sonra hələ də subtle. Bu pattern göstərir ki, **dramatic physics effects** (heavy sag, hard shadows, big splash) model üçün 2-3 iteration tələb edə bilər. Knowledge-da "dramatic physics requires 2+ iterations" qeydi edilir.
|
|
320
|
+
|
|
321
|
+
---
|
|
322
|
+
|
|
323
|
+
## Növbəti test üçün checklist (image-prompt-engineer bu faylı oxuyur)
|
|
324
|
+
|
|
325
|
+
1. **Bu prompt-da subjective quantifier var?** ("realistic", "natural", "dramatic", "soft", "intense") → numeric + anti-default + verify ilə əvəz et
|
|
326
|
+
2. **Bu prompt-da dual/multi-element var?** (multiple banisters, lights, supports, characters) → numbered subsystems + "ALL must" + verification checklist
|
|
327
|
+
3. **Bu prompt-da time-of-day var?** → explicit numeric time + sun position + shadow geometry + color temp + anti-default
|
|
328
|
+
4. **Bu prompt-da physical loading var?** (rope under weight, beam under mass) → numeric sag depth + comparative differential
|
|
329
|
+
5. **Bu prompt-da parantezdə vacib instruction var?** → parantezdən çıxar, ayrı numbered cümlə kimi qoy
|
|
330
|
+
|
|
331
|
+
---
|
|
332
|
+
|
|
333
|
+
*Versiya: 1.0 | Test sayı: 3 (P-01, P-02, P-03) | Son yenilənmə: 2026-05-15*
|
|
334
|
+
|
|
335
|
+
---
|
|
336
|
+
|
|
337
|
+
## 2026-05-15 — P-06 iter 1 → 2 (Rainy Night Street) — Partial fix, model limitations discovered
|
|
338
|
+
|
|
339
|
+
### Test context
|
|
340
|
+
P-06 v1: 1-shot success with 2 ⚠️ (rain angle, pant wet visibility). Yeni qaydaya görə (CLAUDE.md Image-validator error loop — ⚠️ də auto-fix tetikləyir) iter 2 yazıldı.
|
|
341
|
+
|
|
342
|
+
### v1 → v2 nəticə
|
|
343
|
+
- ✅ **Shoe wet sheen ⚠️ → ✅** — "CLEAR SPECULAR HIGHLIGHTS" + "splash marks visible on shoe tops" instruction-larıyla əldə olundu
|
|
344
|
+
- ⚠️ **Rain diagonal angle (15-20°)** — 2 iterasiyada da fix olunmadı. Numeric, anti-default, verify clue heç biri tam işləmədi
|
|
345
|
+
- ⚠️ **Pant knee-line wet contrast** — 2 iterasiyada da dark fabric + dark scene vizual contrast vermədi
|
|
346
|
+
|
|
347
|
+
### Knowledge gap kəşfi — model limitations vs prompt failures
|
|
348
|
+
|
|
349
|
+
Bu 2 iterasiyalı failure göstərdi ki, bəzi ⚠️-lar **model limitation**-dur (prompt failure deyil).
|
|
350
|
+
|
|
351
|
+
#### Limitation 1 — Strong diagonal rain bias (GPT-Image-2)
|
|
352
|
+
- Training data bias: rain images mostly vertical/soft angle
|
|
353
|
+
- Prompt remediation **zəif işləyir**
|
|
354
|
+
- **Tövsiyə:** Strong diagonal rain üçün **Flux Kontext** və ya **Midjourney v7** (--style raw + storm reference) alternative model
|
|
355
|
+
|
|
356
|
+
#### Limitation 2 — Dark fabric wet/dry contrast in low-light
|
|
357
|
+
- Dark fabric specular response intrinsic olaraq zəifdir
|
|
358
|
+
- Wet/dry differential yalnız light-coloured fabric-də aydın render olunur
|
|
359
|
+
- **Tövsiyə (scene preconditioning):**
|
|
360
|
+
- Wet/dry contrast vacibdirsə → **lighter colour fabric** seç (beige, light grey, tan)
|
|
361
|
+
- Və ya wet area üzərinə **bright light source** yerləşdir
|
|
362
|
+
|
|
363
|
+
### Yeni meta-rule: 2-iteration model limitation flag
|
|
364
|
+
|
|
365
|
+
P-06 göstərdi ki, **eyni ⚠️ 2 iterasiyada təkrar olunarsa**, prompt-engineer-in 3-cü iterasiya cəhdi əvəzinə:
|
|
366
|
+
1. **Model limitation flag** qaldırır
|
|
367
|
+
2. **Alternative model** təklif edir (rain → Flux Kontext, fabric contrast → lighter colour)
|
|
368
|
+
3. **Scene preconditioning** tövsiyəsi verir (məs. "light beige trousers seç wet-dry contrast üçün")
|
|
369
|
+
4. **İstifadəçi qərar verir:** accept ⚠️, change model, change scene element
|
|
370
|
+
|
|
371
|
+
### Knowledge update tövsiyəsi
|
|
372
|
+
- [x] Test-learnings.md-ə bu entry əlavə olundu
|
|
373
|
+
- [ ] CLAUDE.md "Image-validator error loop qaydası"-na **"2-iteration model limitation flag"** əlavə olunsun
|
|
374
|
+
- [ ] prompt-consistency-checks.md-ə **F11 — Model intrinsic limitations** əlavə olunsun (rain bias, fabric contrast, və s.)
|
|
375
|
+
- [ ] physical-realism.md bölmə 4 (weather) — "Diagonal rain rendering limitation across image models" qeyd olunsun
|
|
376
|
+
|
|
377
|
+
### Cinema reference pattern — F12 candidate
|
|
378
|
+
P-05 (Sofia Coppola), P-06 v1, P-06 v2 — 3 test-də cinema/director reference həlledici rol oynadı:
|
|
379
|
+
- Lost in Translation aesthetic → Fresnel layered reflection işlədi
|
|
380
|
+
- Blade Runner 2049 / Sicario → wet street long warm reflection mükəmməl
|
|
381
|
+
|
|
382
|
+
**3-test threshold çatdı** → master knowledge-ə promote edilməli pattern:
|
|
383
|
+
|
|
384
|
+
> **F12 — Cinema reference as fix tool (universal pattern):**
|
|
385
|
+
> Kompleks vizual situations üçün uyğun film/director reference promptə əlavə et. Model film aesthetic-i statistical olaraq güclü association saxlayır (training data-da film imagery dominant). Reference müəyyən mood + lighting + composition + physics-i bir paket kimi modelə təqdim edir.
|
|
386
|
+
|
|
387
|
+
---
|
|
388
|
+
|
|
389
|
+
*Versiya: 1.1 | Test sayı: 6 (P-01 → P-06) | Yeni pattern: F10, F11 candidate, F12 candidate | Son yenilənmə: 2026-05-15*
|
|
390
|
+
|
|
391
|
+
---
|
|
392
|
+
|
|
393
|
+
## 2026-05-15 — P-07 iter 1 → 2 (Bakı Caspian Sunset) — Research-first SUCCESS
|
|
394
|
+
|
|
395
|
+
### Test context
|
|
396
|
+
**Test edir:** Geographic accuracy + cultural/landmark specifics + cinema reference
|
|
397
|
+
**Risk pattern:** AI generic stock seaside imagery — real Baku landmarks tanınmır, default geographic biases ("sun into water", "towers at sea level") qarşıya çıxır
|
|
398
|
+
|
|
399
|
+
### Failure (v1)
|
|
400
|
+
**Prompt:** P-07 v1 — "Flame Towers visible on left side, illuminated"
|
|
401
|
+
**Faktiki nəticə:**
|
|
402
|
+
- Flame Towers **boulevard-da sea level-də** rendered olundu (real-life-də 15-dəq pilləkən climbing yuxarıda təpədə)
|
|
403
|
+
- Generic skyline arasında (real Baku architectural mix yox idi)
|
|
404
|
+
- Scale relationship səhv (towers boulevard-a çox yaxın)
|
|
405
|
+
|
|
406
|
+
### Root cause
|
|
407
|
+
- **AI generic real-world location bias:** Model "seaside boulevard + flame-shaped towers" prompt-ı qəbul edib, lakin **konkret real Baku geography**-ni tanımır
|
|
408
|
+
- Training data: çoxsaylı seaside images, lakin Baku-specific imagery az
|
|
409
|
+
- Without research-backed prompt → model default "generic" reproduces
|
|
410
|
+
|
|
411
|
+
### Yeni qayda kəşfi: Geographic/Landmark accuracy research
|
|
412
|
+
|
|
413
|
+
**Bu test, yepyeni sistem qaydasını doğurdu** — CLAUDE.md-də sərt qayda olaraq yazıldı:
|
|
414
|
+
|
|
415
|
+
> Hər dəfə bir konkret real dünya məkanı (şəhər, rayon, küçə, landmark, ünvan) prompt-da göstərilirsə, **əvvəlcə internet araşdırması məcburidir**. WebSearch / WebFetch ilə faktiki məlumat tap, `03-locations/<məkan>-research.md` faylına yaz, prompt yazılarkən bu faktları istifadə et.
|
|
416
|
+
|
|
417
|
+
### Fix tətbiqi (iter 2)
|
|
418
|
+
|
|
419
|
+
**Research nəticələri:**
|
|
420
|
+
- Flame Towers: Highland Park hill, 15-min stair climb up from boulevard, ELEVATED, inland setback
|
|
421
|
+
- Baku Bay: opens south, city on western half, downtown slopes down to sea
|
|
422
|
+
- Sun at sunset: west (behind city/towers from boulevard POV)
|
|
423
|
+
- Caspian: east-southeast of Baku
|
|
424
|
+
|
|
425
|
+
**Prompt rewrite stratejisi:**
|
|
426
|
+
1. Research-based geographic facts (Flame Towers elevation, setback, scale)
|
|
427
|
+
2. Multiple anti-default clauses (NOT at sea level, NOT sun over water)
|
|
428
|
+
3. Layered scene composition (boulevard → city → hill → towers)
|
|
429
|
+
4. Authentic architectural reference (Baku 19th century classical mix)
|
|
430
|
+
5. Cinema reference: "The Cut" (Fatih Akın) — Caspian cinematography
|
|
431
|
+
|
|
432
|
+
### Nəticə
|
|
433
|
+
**v2 = 33/33 ✅ (perfect success).** Bütün ❌-lar düzəldi:
|
|
434
|
+
- ✅ Flame Towers elevated on hill
|
|
435
|
+
- ✅ Authentic Baku architectural mix (classical 19th century buildings real)
|
|
436
|
+
- ✅ Sun west / shadows east
|
|
437
|
+
- ✅ Proper scale relationship
|
|
438
|
+
|
|
439
|
+
### Carry-forward rule
|
|
440
|
+
**For any real-world location/landmark prompt:**
|
|
441
|
+
1. **Research first** — WebSearch faktları tap
|
|
442
|
+
2. **Document** — `03-locations/<məkan>-research.md` faylı yarat
|
|
443
|
+
3. **Prompt-da fakt istifadə et** — generic təxminə yer yox
|
|
444
|
+
4. **Anti-default ultra** — AI generic stock biases-i explicit qır
|
|
445
|
+
5. **Validate against research** — Qat B-də faktiki məlumatla tutuşdur
|
|
446
|
+
|
|
447
|
+
### Knowledge update tövsiyəsi
|
|
448
|
+
- [x] CLAUDE.md-də yeni "Geographic/Landmark accuracy research qaydası" əlavə olundu
|
|
449
|
+
- [x] location-designer/SKILL.md-də Addım 0c (geographic research) əlavə olundu
|
|
450
|
+
- [x] image-prompt-engineer/SKILL.md-də Addım 0-GEO əlavə olundu
|
|
451
|
+
- [x] image-validator script-consistency-checks.md-də Geographic accuracy check əlavə olundu
|
|
452
|
+
- [x] Test research file template `baku-boulevard-research.md` ilə nümayiş olundu
|
|
453
|
+
|
|
454
|
+
### Sistem dəyəri (P-07 = paradigm shift)
|
|
455
|
+
|
|
456
|
+
P-07 öyrətdi ki, sistemə yalnız **prompt-image consistency** yetərli deyil — **prompt-reality consistency** də vacibdir. Real-world referansları olan layihələr (məs. Baku Reels test) artıq research-first workflow-da işləyəcək. Bu, sistemin **professional production-grade** səviyyəyə ucalması üçün kritik qayda idi.
|
|
457
|
+
|
|
458
|
+
---
|
|
459
|
+
|
|
460
|
+
*Versiya: 1.2 | Test sayı: 7 (P-01 → P-07) | Yeni qaydalar: F10, F11 candidate, F12 candidate, **Geographic research rule** | Son yenilənmə: 2026-05-15*
|
|
461
|
+
|
|
462
|
+
---
|
|
463
|
+
|
|
464
|
+
## 2026-05-15 — P-08 (Pouring Tea — Dynamic Material Physics) — SUCCESS with 1 ⚠️ → fix prompt yazıldı
|
|
465
|
+
|
|
466
|
+
### Test context
|
|
467
|
+
**Test edir:** Material physics dynamic — water arc + splash + bubbles + ripples + glass refraction + steam
|
|
468
|
+
**Risk pattern:** AI dynamic water physics tipik weak point — stream sıxlıqla straight vertical, no splash, opaque glass
|
|
469
|
+
|
|
470
|
+
### Pattern application (proactive 7 patterns)
|
|
471
|
+
1. F8 numeric specificity (stream "15-18cm length, 5mm→3mm narrowing")
|
|
472
|
+
2. F9 dual-element (7 numbered subsystems)
|
|
473
|
+
3. F10 Fresnel (glass refraction + tea surface ripples + saucer glow)
|
|
474
|
+
4. F11 watch (dynamic water known weak spot)
|
|
475
|
+
5. F12 cinema reference ("Chef's Table" Netflix + Penny De Los Santos)
|
|
476
|
+
6. Anti-default ULTRA (6 NOT-clauses)
|
|
477
|
+
7. Verification checklist (9 items + 4 failure conditions)
|
|
478
|
+
|
|
479
|
+
### Nəticə
|
|
480
|
+
**24 ✅, 2 ⚠️, 0 ❌** — first-shot success
|
|
481
|
+
- ⚠️ Stream curve subtle (diagonal-dominant, parabolic not clear)
|
|
482
|
+
- ✅ Crown splash with 8+ droplets — STRONG
|
|
483
|
+
- ✅ Glass refraction — STRONG (tea inside cup distorted)
|
|
484
|
+
- ✅ Amber glow on saucer — STRONG
|
|
485
|
+
- ✅ Bubbles + ripples + steam — all rendered
|
|
486
|
+
|
|
487
|
+
### Bonus: Cultural authenticity
|
|
488
|
+
- Cup rendered as Azerbaijani armudu (pear-shaped) glass — culturally authentic without explicit prompt
|
|
489
|
+
- Background plate with dessert — authentic tea service detail
|
|
490
|
+
- Hypothesis: cinema reference "Chef's Table" + tea pouring + warm light context may have biased model toward "Caucasus/Turkish/Azerbaijani tea aesthetic" — desirable side effect
|
|
491
|
+
|
|
492
|
+
### Fix iteration 1 yazıldı (⚠️ auto-fix qaydası tətbiqi)
|
|
493
|
+
Stream curve ⚠️ üçün aqressiv prompt yenidən yazıldı:
|
|
494
|
+
- Numeric angle progression: 45° (top) → 70° (mid) → 85° (impact)
|
|
495
|
+
- "Thrown projectile trajectory" reference
|
|
496
|
+
- "Parabolic arc" explicit language
|
|
497
|
+
- "Vertical pole vs mid-arc ~25° angle" verify clue
|
|
498
|
+
- Failure condition: "if stream looks straight diagonal → FAILURE"
|
|
499
|
+
|
|
500
|
+
### Carry-forward rule (water stream physics)
|
|
501
|
+
**For any "liquid pour / falling water" prompt:**
|
|
502
|
+
1. Numeric angle progression (start → mid → end)
|
|
503
|
+
2. "Parabolic arc" + "thrown projectile" reference language
|
|
504
|
+
3. Width narrowing numeric (mm at each point)
|
|
505
|
+
4. Anti-default: "NOT straight diagonal", "NOT vertical column"
|
|
506
|
+
5. Verify clue: angle comparison to a reference vertical element
|
|
507
|
+
|
|
508
|
+
### Knowledge update tövsiyəsi
|
|
509
|
+
- [x] Test-learnings.md-ə bu entry əlavə olundu
|
|
510
|
+
- [ ] physical-realism.md bölmə 6.1 (Material — su) — "Falling water parabolic arc rule" əlavə olunsun (Discovered P-08)
|
|
511
|
+
|
|
512
|
+
### Pattern statistics — F9 universal success
|
|
513
|
+
P-04 (handshake 2 hands), P-05 (window 2 layers), P-06 (umbrella zones), P-08 (7 fluid components) — F9 numbered subsystem rule **4 test-də uğurla işlədi**. Sübut olunmuş universal pattern.
|
|
514
|
+
|
|
515
|
+
---
|
|
516
|
+
|
|
517
|
+
*Versiya: 1.3 | Test sayı: 8 (P-01 → P-08) | Yeni qaydalar: F10, F11, F12, Geographic research, Parabolic arc | Son yenilənmə: 2026-05-15*
|
|
518
|
+
|
|
519
|
+
---
|
|
520
|
+
|
|
521
|
+
## 2026-05-15 — P-08 v2 (Armudu Glass Geometry) — Cultural Object Research kəşfi
|
|
522
|
+
|
|
523
|
+
### Failure (v2)
|
|
524
|
+
**Problem:** Armudu istəkan səhv rendered olundu — tapered tulip kimi (yuxarı enli, aşağı dar). Real armudu **ƏKSİNƏ**: bottom bulb enli, narrow waist ortada, top dar.
|
|
525
|
+
**❌ Pozulma kateqoriyası:** Cultural object accuracy + AI generic "tea glass" bias
|
|
526
|
+
|
|
527
|
+
### Root cause
|
|
528
|
+
- AI training data-da "tea glass" çox yayılmış generic tulip/tapered shape kimi mövcuddur
|
|
529
|
+
- Armudu specific shape under-represented — model fərq tanımır
|
|
530
|
+
- Prompt-da "armudu glass" yazılsa da, model **shape-i öyrətmir** sadəcə "tea-related glass" associativ assosasiya edir
|
|
531
|
+
- Generic "tapered" word triggered standard tulip render
|
|
532
|
+
|
|
533
|
+
### Yeni qayda kəşfi: Cultural Object Research
|
|
534
|
+
|
|
535
|
+
**Bu test Geographic accuracy research qaydasını genişləndirdi** — yalnız geographic location-lar deyil, **konkret mədəni obyektlər** də araşdırma tələb edir:
|
|
536
|
+
|
|
537
|
+
CLAUDE.md-də Geographic accuracy research qaydasının trigger cədvəlinə əlavə olundu:
|
|
538
|
+
- Cultural objects (armudu, kimono, samovar, matryoshka, traditional clothing)
|
|
539
|
+
- Cultural food (Azerbaijani plov, ramen, paella — specific presentation)
|
|
540
|
+
- Musical instruments (tar, sitar, balalaika)
|
|
541
|
+
|
|
542
|
+
### Fix tətbiqi (iter 2 → 3)
|
|
543
|
+
|
|
544
|
+
**Research nəticələri (Wikipedia + Azerbaijani sources):**
|
|
545
|
+
- Armudu: pear-shaped, "Boğmalı" = narrow-waisted
|
|
546
|
+
- Geometry yuxarıdan aşağıya:
|
|
547
|
+
- Rim ~5cm
|
|
548
|
+
- Upper bell narrowing
|
|
549
|
+
- **Waist ~3.5cm** (narrowest)
|
|
550
|
+
- **Bottom bulb ~6.5cm** (WIDEST — counter-intuitive)
|
|
551
|
+
- Foot ~3.5cm
|
|
552
|
+
- Bottom bulb diameter > top rim diameter — **defining feature**
|
|
553
|
+
- Volume 100-150ml, height 8cm
|
|
554
|
+
|
|
555
|
+
**Prompt rewrite stratejisi (v3 FULL prompt, kopya-hazır):**
|
|
556
|
+
1. Section 4 (4A-4I) — 9 alt-bölmə armudu geometry
|
|
557
|
+
2. Female silhouette / pear analogy
|
|
558
|
+
3. Numeric proportions (waist 3.5cm, bottom 6.5cm)
|
|
559
|
+
4. 6 anti-default clauses (NOT tulip, NOT snifter, NOT cylinder, NOT wine, NOT hourglass, NOT tapered)
|
|
560
|
+
5. "Verify: is bottom bulb wider than top rim?" verification
|
|
561
|
+
6. Failure condition: "If NOT armudu shape → regenerate"
|
|
562
|
+
|
|
563
|
+
### Carry-forward rule (cultural objects)
|
|
564
|
+
**For any culturally-specific object in prompt:**
|
|
565
|
+
1. WebSearch faktiki geometry/proportions/appearance
|
|
566
|
+
2. Document `03-locations/<object>-research.md` faylına
|
|
567
|
+
3. Prompt-da numeric proportions + analogies + anti-defaults
|
|
568
|
+
4. Verification specific to that object's defining features
|
|
569
|
+
5. Failure conditions explicit
|
|
570
|
+
|
|
571
|
+
### Knowledge update tövsiyəsi
|
|
572
|
+
- [x] Test-learnings.md-ə bu entry əlavə olundu
|
|
573
|
+
- [x] CLAUDE.md "Geographic accuracy research" trigger cədvəlinə cultural objects əlavə olundu
|
|
574
|
+
- [x] CLAUDE.md "Niyə bu qayda var" bölməsinə P-08 v2 cultural extension qeyd olundu
|
|
575
|
+
- [x] `03-locations/armudu-glass-research.md` research file yaradıldı
|
|
576
|
+
|
|
577
|
+
### Meta-rule: Cultural specificity is geographic accuracy
|
|
578
|
+
P-07 (Bakı Flame Towers) və P-08 v2 (armudu glass) eyni meta-pattern göstərdi: **AI generic stock representations öyrənir, cultural-specific obyektlər under-represented olur**. Research-first yanaşma hər iki halda kritik. Sistem indi tanıyır ki, "Azerbaijani context" trigger olunan layihələrdə **bütün cultural elements** research file tələb edir.
|
|
579
|
+
|
|
580
|
+
### Yeni sərt qayda: Fix mode FULL PROMPT məcburi (P-08 fix-də pozulma)
|
|
581
|
+
Bu iteration-da mən səhv etdim — P-08 v2 fix-i fragment kimi yazdım ("eyni v1 prompt + dəyişən section"). İstifadəçi kopyalaya bilmədi. CLAUDE.md + image-prompt-engineer SKILL.md-də sərt qayda olaraq yazıldı: **Fix mode-da FULL PROMPT məcburi, fragment qadağan**.
|
|
582
|
+
|
|
583
|
+
P-08 v3 bu qaydaya uyğun yazıldı — başdan-sona tam prompt, kopya-paste hazır.
|
|
584
|
+
|
|
585
|
+
---
|
|
586
|
+
|
|
587
|
+
*Versiya: 1.4 | Test sayı: 8 (P-01 → P-08 — 3 iteration) | Yeni qaydalar: Cultural object research + Fix mode FULL PROMPT | Son yenilənmə: 2026-05-15*
|
|
588
|
+
|
|
589
|
+
---
|
|
590
|
+
|
|
591
|
+
## 2026-05-15 — P-08 v3 (Armudu Geometry + Parabolic Arc) — Dramatic Success
|
|
592
|
+
|
|
593
|
+
### v2 → v3 nəticə
|
|
594
|
+
**31/32 ✅** (1 minor ⚠️):
|
|
595
|
+
- ⚠️ Stream subtle curve (v2) → ✅ Dramatic parabolic arc (v3) — **PERFECT**
|
|
596
|
+
- ❌ Wrong glass shape tapered tulip (v2) → ✅ Authentic armudu pear-shape (v3) — **PERFECT**
|
|
597
|
+
- BONUS: Buta pattern + decorative saucer + sugar cubes bowl + authentic Azerbaijani tea ceremony aesthetic
|
|
598
|
+
|
|
599
|
+
Remaining ⚠️: AR-7 — Bottom bulb vs top rim ratio close to equal, real armudu has bottom slightly more pronounced.
|
|
600
|
+
|
|
601
|
+
### Cultural Object Research qaydasının ilk uğurlu real tətbiqi
|
|
602
|
+
P-08 v2→v3 göstərdi ki, P-07 geographic accuracy ilə paralel — **Cultural Object Research** universal pattern:
|
|
603
|
+
- Generic AI training → wrong cultural object
|
|
604
|
+
- Research-first → authentic representation
|
|
605
|
+
|
|
606
|
+
İki uğurlu real tətbiq (P-07 + P-08 v3) → sistemin **professional production-grade** səviyyəsini sübut edir.
|
|
607
|
+
|
|
608
|
+
### Fix mode FULL PROMPT qaydası — ilk uğurlu tətbiq
|
|
609
|
+
P-08 v3 sənin yeni qaydan altında ilk **FULL PROMPT** Fix iteration-dur:
|
|
610
|
+
- v2 prompt-un bütün bölmələri v3-də saxlandı
|
|
611
|
+
- ❌ olan sections (stream + glass) aqressiv yenidən yazıldı
|
|
612
|
+
- ✅ olan sections olduğu kimi qaldı
|
|
613
|
+
- İstifadəçi tam kopya-paste edə bildi
|
|
614
|
+
|
|
615
|
+
İstifadəçi tempo qənaət etdi, sistem doğru davrandı.
|
|
616
|
+
|
|
617
|
+
### Meta-pattern: 2 research-first qaydası eyni vaxtda tətbiq olundu
|
|
618
|
+
Bu test (P-08 v3) həm **physical-realism research** (parabolic arc fluid dynamics) həm **cultural object research** (armudu geometry) eyni promptda kombinasiya etdi. Both worked. Bu, **multi-research workflow**-un işlədiyini sübut edir.
|
|
619
|
+
|
|
620
|
+
### Knowledge update tövsiyəsi
|
|
621
|
+
- [x] Test-learnings.md-ə bu v3 success entry əlavə olundu
|
|
622
|
+
- [ ] P-08 üzərindən "Fluid dynamics + cultural object research" combined approach physical-realism.md-də cross-reference olaraq qeyd olunsun
|
|
623
|
+
|
|
624
|
+
---
|
|
625
|
+
|
|
626
|
+
*Versiya: 1.5 | Test sayı: 8 (P-01 → P-08, 3 iterations) | Universal patterns confirmed: F9 (4x), F12 (3x), Cultural research (2x), FULL PROMPT (1x) | Son yenilənmə: 2026-05-15*
|
|
627
|
+
|
|
628
|
+
---
|
|
629
|
+
|
|
630
|
+
## 2026-05-15 — P-08 v4 (Armudu Ratio Perfection) — Full Cultural Success
|
|
631
|
+
|
|
632
|
+
### v3 → v4 nəticə
|
|
633
|
+
**32/32 ✅** — heç bir ⚠️ və ya ❌ qalmadı.
|
|
634
|
+
|
|
635
|
+
### Hansı texnikalar həlledici işlədi (v3-dən v4-yə):
|
|
636
|
+
1. **Explicit numeric ratio:** "~1.5x ratio bottom:top"
|
|
637
|
+
2. **Pomegranate analogy** (cultural symbol) — "narrow top, rounded full bottom"
|
|
638
|
+
3. **Visual verification:** "the bottom should LOOK CLEARLY BIGGER than the top"
|
|
639
|
+
4. **Anti-default genişləndirildi:** "NOT a glass with bottom bulb ≈ top rim"
|
|
640
|
+
5. **Failure condition explicit:** "If bottom ≈ top → FAILURE → regenerate"
|
|
641
|
+
6. **Smaller numeric for top rim:** ~4.5cm (v3-də ~5cm idi)
|
|
642
|
+
7. **Larger numeric for bulb:** ~7cm (v3-də ~6.5cm idi)
|
|
643
|
+
8. **Smaller numeric for waist:** ~3cm (v3-də ~3.5cm idi)
|
|
644
|
+
|
|
645
|
+
### Bonus emergent properties (model interpreted "authentic" liberally)
|
|
646
|
+
v4 promptda "embroidered Azerbaijani tablecloth pattern" əlavə olundu, model bunu **traditional carpet pattern** kimi render etdi. Bu, **emergent cultural authenticity** göstərir — model "Azerbaijani tea ceremony" context-da kompleks **multiple authentic elements** kombinasiya edir:
|
|
647
|
+
- Buta paisley on glass walls (intricate)
|
|
648
|
+
- Traditional carpet under setting
|
|
649
|
+
- Decorative gold-red saucer
|
|
650
|
+
- Sugar cube bowl
|
|
651
|
+
- Ornate samovar-style teapot
|
|
652
|
+
- Window backlight home atmosphere
|
|
653
|
+
|
|
654
|
+
Bu, P-04 (handshake-də Azerbaijani context emerging) ilə paralel pattern. Hypothesis: **strong cultural context activation** → model **catalogue of authentic elements** çağırır.
|
|
655
|
+
|
|
656
|
+
### Pattern: Numeric ratios > qualitative descriptions
|
|
657
|
+
P-08 v3-də "wider than top rim" yetərli deyildi (model interpreted as "slightly wider, similar"). v4-də **explicit 1.5x ratio** + **visual verification** + **failure condition** trio həlledici idi.
|
|
658
|
+
|
|
659
|
+
**Carry-forward rule:**
|
|
660
|
+
> Qualitative comparative ("wider than", "larger", "more") prompt-da yetərli deyil. Numeric ratio + explicit visual check + failure condition trio tələb olunur.
|
|
661
|
+
|
|
662
|
+
### Knowledge update tövsiyəsi
|
|
663
|
+
- [x] Test-learnings.md-ə bu v4 PERFECT success entry əlavə olundu
|
|
664
|
+
- [ ] prompt-consistency-checks.md-ə yeni pattern "Numeric ratio precision" əlavə olunsun (qualitative comparison-dan numeric ratio-ya yüksəlmə)
|
|
665
|
+
|
|
666
|
+
### Statistical summary (P-08 final, 4 iterations):
|
|
667
|
+
|
|
668
|
+
| Version | Result | Major learning |
|
|
669
|
+
|---|---|---|
|
|
670
|
+
| v1 | ⚠️×2 | F9 + cinema reference proactive worked |
|
|
671
|
+
| v2 | ❌ + ⚠️ | Cultural object research kəşfi (armudu generic-dən fərqli) |
|
|
672
|
+
| v3 | ✅ + ⚠️ | FULL PROMPT qaydası ilk uğurlu tətbiq |
|
|
673
|
+
| **v4** | **✅ Perfect** | **Numeric ratio precision rule** |
|
|
674
|
+
|
|
675
|
+
### Meta-confirmation: 4 iteration P-08-də sübut olundu
|
|
676
|
+
- F9 universal (P-04, P-05, P-06, P-08)
|
|
677
|
+
- F12 cinema reference (P-05, P-06, P-07, P-08)
|
|
678
|
+
- Cultural object research (P-07, P-08)
|
|
679
|
+
- FULL PROMPT (P-08 v3, v4)
|
|
680
|
+
- Numeric ratio precision (P-08 v4 — new pattern)
|
|
681
|
+
|
|
682
|
+
---
|
|
683
|
+
|
|
684
|
+
*Versiya: 1.6 | Test sayı: 8 (P-01 → P-08, 4 iterations cumulative) | Yeni pattern: Numeric ratio precision | Son yenilənmə: 2026-05-15*
|
|
685
|
+
|
|
686
|
+
---
|
|
687
|
+
|
|
688
|
+
## 2026-05-15 — P-09 (Cat + Bird at Window — Multi-animal Anatomy + Fresnel) — 1-SHOT PERFECT SUCCESS
|
|
689
|
+
|
|
690
|
+
### Test context
|
|
691
|
+
**Test edir:** Cat anatomy + bird anatomy + window Fresnel layering simultaneously
|
|
692
|
+
**Risk pattern:** Animal anatomy AI tipik weakness (extra ears, wrong leg count, missing toes), multi-subject coordination
|
|
693
|
+
|
|
694
|
+
### Pattern application (4 proactive patterns kombinasiya)
|
|
695
|
+
1. **F9 dual-element** — Cat (5 subsystems) + Bird (5 subsystems) + Window (3 subsystems) = 13 numbered subsystems
|
|
696
|
+
2. **F10 Fresnel surface** — daytime overcast = ~15% reflection opacity (less than night, but visible)
|
|
697
|
+
3. **F11 animal anatomy precision** — exact counts (2 ears, 2 eyes, 4 legs, 1 tail, 2 wings, 2 legs, 3-4 toes, 1 beak)
|
|
698
|
+
4. **F12 cinema reference** — "The Cat Returns" Ghibli + Walter Chandoha cat photographer
|
|
699
|
+
|
|
700
|
+
### Numeric precision rule tətbiqi
|
|
701
|
+
P-08 v4-dən gələn "Numeric Ratio Precision" pattern bu test-də də işlədi:
|
|
702
|
+
- "EXACTLY 2 ears (not 1, not 3, not 4)" — explicit count
|
|
703
|
+
- "10-12 whiskers per side" — explicit range
|
|
704
|
+
- "3-4 toes per foot" — explicit range
|
|
705
|
+
- Verify clauses: "count toes — should be 3-4 visible per foot"
|
|
706
|
+
|
|
707
|
+
### Nəticə
|
|
708
|
+
**30/30 ✅** — first-shot perfect success. Heç bir ⚠️/❌ yoxdur.
|
|
709
|
+
|
|
710
|
+
### Emergent authenticity (yenidən təkrar olundu — P-08 ilə paralel)
|
|
711
|
+
Model "intimate domestic moment + cat watching bird" context-da **avtomatik** authentic details əlavə etdi:
|
|
712
|
+
- ✨ Realistic dust particles on glass
|
|
713
|
+
- ✨ Spring tree buds (seasonal authenticity)
|
|
714
|
+
- ✨ Worn paint on windowsill
|
|
715
|
+
- ✨ Real Eurasian Tree Sparrow features (correct species)
|
|
716
|
+
- ✨ Cat's rim light from window glow
|
|
717
|
+
|
|
718
|
+
**Hypothesis:** Strong cinema/photographer reference + numbered subsystem prompt = model **expands authentic context** beyond explicit instructions. Bu, **professional production-grade** sıçrayışın əlamətidir.
|
|
719
|
+
|
|
720
|
+
### F11 (animal anatomy) — yeni confirmation
|
|
721
|
+
P-09 göstərdi ki, **multi-animal anatomy** prompt-da:
|
|
722
|
+
1. EXACT count per body part (numeric)
|
|
723
|
+
2. Anti-default 3-clause per body part ("NOT 1, NOT 3, NOT 4")
|
|
724
|
+
3. Verify clause ("count toes — should be 3-4")
|
|
725
|
+
4. Per-animal numbered subsystem (1A-1E, 2A-2E)
|
|
726
|
+
|
|
727
|
+
Kombinasiya işləyir. Heç bir cat-da 3 qulaq, heç bir bird-də 3 qanad rendered olunmadı.
|
|
728
|
+
|
|
729
|
+
### Statistical update (9 test sonra):
|
|
730
|
+
|
|
731
|
+
| Test | Iter | Result | Confirmed patterns |
|
|
732
|
+
|---|---|---|---|
|
|
733
|
+
| P-01 | 2 | ✅ | F8 |
|
|
734
|
+
| P-02 | 2+ | ✅ partial | Drama subjective |
|
|
735
|
+
| P-03 | 2 | ✅ | F9 discovered |
|
|
736
|
+
| P-04 | 1 ✨ | ✅ | Proactive F6+F9 |
|
|
737
|
+
| P-05 | 2 | ✅ | F10 discovered |
|
|
738
|
+
| P-06 | 2 | ✅ partial | F11 model limitation |
|
|
739
|
+
| P-07 | 2 | ✅ | **Geographic research** |
|
|
740
|
+
| P-08 | 4 | ✅ Perfect | **Cultural object research + FULL PROMPT + Numeric Ratio** |
|
|
741
|
+
| **P-09** | **1** ✨ | **✅ Perfect** | **Multi-animal anatomy + Fresnel + 4-pattern kombinasiya** |
|
|
742
|
+
|
|
743
|
+
### Knowledge update tövsiyəsi
|
|
744
|
+
- [x] Test-learnings.md-ə bu entry əlavə olundu
|
|
745
|
+
- [ ] physical-realism.md bölmə 5 (Anatomy) — "Multi-animal exact count pattern" əlavə olunsun
|
|
746
|
+
|
|
747
|
+
---
|
|
748
|
+
|
|
749
|
+
*Versiya: 1.7 | Test sayı: 9 (P-01 → P-09, P-09 1-shot perfect) | Universal patterns confirmed multiple times | Son yenilənmə: 2026-05-15*
|
|
750
|
+
|
|
751
|
+
---
|
|
752
|
+
|
|
753
|
+
## 2026-05-15 — Reels Project Pilləkən v1 → v3 → v4 (Real Production Application)
|
|
754
|
+
|
|
755
|
+
### Test context
|
|
756
|
+
**Production layihəsi:** `/tmp/creator-test-3/girl-and-cat-friendship/`
|
|
757
|
+
**Visual style:** pixar-3d (Disney/Pixar 3D animation)
|
|
758
|
+
**Image model:** GPT-Image-2 (project lock)
|
|
759
|
+
|
|
760
|
+
### v1 → v3 (catastrophic bug fix)
|
|
761
|
+
**v1 problem:** "Yarısı yoxdu" — half stair structure missing into void (P-03-də yazdığımız F9 dual-banister rule yox idi)
|
|
762
|
+
**v3 fix:** Bilateral stair structure explicit — inner wall + outer iron banister, 17 numbered subsystems, numeric ratios, verify checklist
|
|
763
|
+
**Nəticə:** Catastrophic bug → production-grade reference (25 ✅, 2 ⚠️, 0 ❌)
|
|
764
|
+
|
|
765
|
+
### v3 → v4 (auto-fix for ⚠️) — Model Limitation Discovered
|
|
766
|
+
**v3 ⚠️ items:** (1) Pixar style photoreal-leaning, (2) Ceiling bulb subtle glow
|
|
767
|
+
**v4 attempt:** Aggressive Pixar style header + 4 NOT-clauses for bulb + failure conditions
|
|
768
|
+
**v4 nəticə:** **Same 2 ⚠️ persist** — no marginal improvement
|
|
769
|
+
|
|
770
|
+
### 2-iteration model limitation flag tətbiqi (P-06 v2 qaydası)
|
|
771
|
+
v3 + v4 = 2 iteration ilə same ⚠️ = **model intrinsic limitation flag**:
|
|
772
|
+
|
|
773
|
+
**Yeni limitations aşkar olundu (GPT-Image-2):**
|
|
774
|
+
1. **Pixar 3D rendering bias** — model "Pixar 3D" instruction-u qəbul edir lakin photoreal default-a meyl edir. Training data-da photoreal interior dominant. Pixar-specific stylization (idealized proportions, appeal-driven materials) zəif tətbiq olunur.
|
|
775
|
+
2. **Subtle bulb glow persistent** — model bulb fixture-ı "warm interior accent" kimi qəbul edir, 4 anti-default + verify + failure condition belə də güclü override etmir.
|
|
776
|
+
|
|
777
|
+
### Project-level decision logic (yeni meta-pattern)
|
|
778
|
+
**Key insight:** Layihənin **previous refs** (balkon, mətbəx, Ayşən, pişik) də GPT-Image-2-də generasiya olunub və onlar da **eyni photoreal-leaning Pixar style** daşıyır. Yəni:
|
|
779
|
+
|
|
780
|
+
- **Style consistency layihə daxilində qurulub** — bütün refs eyni səviyyədə
|
|
781
|
+
- Pilləkən v3 bu style ilə **uyğundur**
|
|
782
|
+
- Model dəyişmək (Midjourney v7 stronger Pixar) pilləkən üçün daha yaxşı olar, **lakin digər refs ilə inconsistency yaradar**
|
|
783
|
+
|
|
784
|
+
### Carry-forward rule: Project Style Consistency > Individual Image Optimization
|
|
785
|
+
**Real production-da:**
|
|
786
|
+
> Bir layihənin ilk refs müəyyən style səviyyəsində qurulduqdan sonra, sonrakı refs eyni model + eyni style səviyyəsində qalmalıdır. Bir ref-i "daha mükəmməl" stylize etmək consistency-ni pozur.
|
|
787
|
+
|
|
788
|
+
Bu yeni meta-rule **Project Style Lock** anlamını gücləndirir:
|
|
789
|
+
- İlk image model seçimi sonrakı bütün refs üçün **kilidlənir**
|
|
790
|
+
- Model intrinsic limitations layihə üçün **uniform** qəbul edilir
|
|
791
|
+
- "Better quality individual" əvəzinə "consistent project style"
|
|
792
|
+
|
|
793
|
+
### Knowledge update tövsiyəsi
|
|
794
|
+
- [x] Test-learnings.md-ə bu entry əlavə olundu
|
|
795
|
+
- [ ] CLAUDE.md "Model lock qaydası"na "Project style consistency > individual optimization" alt-bölmə əlavə olunsun
|
|
796
|
+
- [ ] GPT-Image-2 known limitations list yaradılsın (Pixar style bias, bulb glow persistence — gələcək layihələrdə avtomatik məlumat)
|
|
797
|
+
|
|
798
|
+
### Statistical update (10 test, real production qoşuldu):
|
|
799
|
+
|
|
800
|
+
| Test | Iter | Result | Knowledge |
|
|
801
|
+
|---|---|---|---|
|
|
802
|
+
| P-01 → P-09 | various | Stress test patterns | F8, F9, F10, F11, F12, Cultural research, FULL PROMPT, Numeric Ratio |
|
|
803
|
+
| **Reels Pilləkən** | **v1 → v3 → v4** | **v3 accepted, model limitations flagged** | **Project Style Lock + 2-iter model limitation real-world tətbiqi** |
|
|
804
|
+
|
|
805
|
+
### Bu test-in dəyəri (production-grade signal)
|
|
806
|
+
Reels pilləkən v1→v3→v4 göstərdi ki:
|
|
807
|
+
1. ✅ **Bilateral stair rule (P-03 kəşfi) real production-da işlədi** — catastrophic v1 bug fixed
|
|
808
|
+
2. ✅ **17 numbered subsystem real prompt-da scalable** (stress test-də 7-13 idi, burada daha böyük)
|
|
809
|
+
3. ✅ **2-iteration model limitation flag pratiq tətbiq olundu** — sistem "wisdom of restraint" göstərdi
|
|
810
|
+
4. ✅ **Project style consistency reasoning aşkar oldu** — individual perfection-dan layihə uniform-luğuna keçid
|
|
811
|
+
|
|
812
|
+
Sistem indi production-da real layihələrlə işləməyə hazırdır.
|
|
813
|
+
|
|
814
|
+
---
|
|
815
|
+
|
|
816
|
+
*Versiya: 1.8 | Test sayı: 9 stress + 1 production (Reels pilləkən) | Yeni meta-rule: Project Style Lock | Son yenilənmə: 2026-05-15*
|
|
817
|
+
|
|
818
|
+
---
|
|
819
|
+
|
|
820
|
+
## Entry: Reels Cell 2 v1 (Production Cell — Beat 1 Ayşən sad MCU)
|
|
821
|
+
|
|
822
|
+
**Tarix:** 2026-05-15
|
|
823
|
+
**Test növü:** Real production (Reels — Qız uşağı ilə pişiyin dostluğu)
|
|
824
|
+
**Faza:** Phase 2 production cells
|
|
825
|
+
**Iteration:** v1 — **FAILED 3 critical issues**
|
|
826
|
+
|
|
827
|
+
### Failure description (3 problem)
|
|
828
|
+
|
|
829
|
+
**Problem 1: Identity mismatch (KRİTİK ❌)**
|
|
830
|
+
- Reference (Ayşən-ref.png): Ayşən üzü konkret — yumru yanaqlar, böyük yaşıl gözlər, dalğalı qara saç, olive-tan dəri, child Pixar appeal
|
|
831
|
+
- Output: Generic Pixar girl — fərqli üz forması, fərqli yanaq nisbəti, fərqli göz formatı, fərqli saç stili
|
|
832
|
+
- **Səbəb:** Reference attachment format yanlış idi — `@img1 = aysen-ref.png` syntax modelə **identity bind yaratmadı**
|
|
833
|
+
|
|
834
|
+
**Problem 2: Cat presence in Cell 2 (KRİTİK ❌)**
|
|
835
|
+
- Cell 2 = Beat 1 ssenari: "AYŞƏN tək oturur, kədərli, yaşlar yoxdur." — pişik OLMAMALI
|
|
836
|
+
- Cell 3 = Beat 2: pişik ilk dəfə həyətdə görünür (Ayşən onu görür)
|
|
837
|
+
- Output: Cell 2-də pişik kadrın aşağı-sağ küncündə (üstəlik perspektiv-də çox böyük)
|
|
838
|
+
- **Səbəb:** Mən `04-heyet-ref.png`-i "color continuity üçün optional" attach etdim → model həyət ref-dən pişiyi götürdü və Cell 2-yə render etdi
|
|
839
|
+
|
|
840
|
+
**Problem 3: AZ text description-da technical jargon (sərt qayda pozuldu)**
|
|
841
|
+
- Mənim yazdığım: "Kompozisiya: MCU, 3/4 right side angle. Üzü rule-of-thirds intersection-da, 50mm equivalent, f/2.8 shallow DOF"
|
|
842
|
+
- İstifadəçi izahı: "Şəkil haqqında təsvir yazanda mütləq geniş və başa düşülən dildə yaz texniki dildə yox. Bu sərt qaydadır"
|
|
843
|
+
- **Səbəb:** Mətn təsvirin məqsədi istifadəçinin anlamasıdır — kinematik termin lüğəti deyil. İstifadəçi yoxlamaq üçün vaxt itirir.
|
|
844
|
+
|
|
845
|
+
### Root cause analysis
|
|
846
|
+
|
|
847
|
+
#### F13 — Reference attachment tag syntax doesn't bind identity
|
|
848
|
+
**New failure pattern aşkar olundu:**
|
|
849
|
+
|
|
850
|
+
Test-də göstərildi ki, prompt mətnində `"image 1 = aysen-ref.png"` və ya `"reference image 1: Ayşən"` syntax **GPT-Image-2-yə identity-ni etibarlı şəkildə bağlamır**. Model:
|
|
851
|
+
- Faylın adını görmür (yalnız upload sırasını bilir)
|
|
852
|
+
- "image 1" tag-ı sadəcə string kimi qəbul edir, "preserve identity" siqnalı oxumur
|
|
853
|
+
- Generic Pixar girl render edir, attached ref-i style template kimi istifadə edir
|
|
854
|
+
|
|
855
|
+
**Doğru pattern (yeni kəşf):**
|
|
856
|
+
- "PRIMARY IDENTITY LOCK: The girl in the FIRST UPLOADED REFERENCE PHOTO is X. Preserve her face EXACTLY..."
|
|
857
|
+
- + specific element copy: "same eye shape, same eye color, same cheek proportions, same hair texture and color, same skin tone"
|
|
858
|
+
- + verify clause: "If your output's face does not match the reference's face, regenerate."
|
|
859
|
+
|
|
860
|
+
Bu format model-də **identity binding mechanism**-i tetikləyir.
|
|
861
|
+
|
|
862
|
+
#### F14 — Beat boundary leakage via optional reference attach
|
|
863
|
+
**New failure pattern:**
|
|
864
|
+
|
|
865
|
+
"Color continuity üçün optional attach" → model attach olunan hər elementi **aktiv** istifadə edir. Cell 2-də həyət ref attach olundu → həyət ref-də pişik var idi → model pişiyi Cell 2 kadrına render etdi. Bu beat boundary pozulmasıdır.
|
|
866
|
+
|
|
867
|
+
**Anti-pattern (qadağan):**
|
|
868
|
+
- "Image 3 (optional): heyət ref — for color continuity"
|
|
869
|
+
- "Bonus reference for atmosphere"
|
|
870
|
+
- "Context image (not for direct use)"
|
|
871
|
+
|
|
872
|
+
**Doğru pattern:**
|
|
873
|
+
- Yalnız o cell-in beat-ində iştirak edən elementlər attach olunur
|
|
874
|
+
- Color continuity prompt mətnində təsvir olunur, attach yox
|
|
875
|
+
|
|
876
|
+
### Fix strategiyası (Cell 2 v2 üçün)
|
|
877
|
+
|
|
878
|
+
1. **Attach minimal:** Yalnız Ayşən ref + Balkon ref (heyət ref ÇIXARILIR)
|
|
879
|
+
2. **Identity lock language** (Primary identity lock + verify clause)
|
|
880
|
+
3. **Beat-strict prompt:** Heç bir pişik / həyət element prompt-da yox
|
|
881
|
+
4. **AZ description plain language:** texniki kinematik termin çıxarılır, plain AZ paragraph yazılır
|
|
882
|
+
|
|
883
|
+
### Carry-forward rules
|
|
884
|
+
|
|
885
|
+
**F13 carry-forward (universal):**
|
|
886
|
+
> Reference attach edəndə, prompt mətnində identity lock dili **məcburi**: "FIRST UPLOADED IMAGE shows X, preserve face EXACTLY" + specific elements + verify clause. `@img1 = filename.png` formatı QADAĞAN.
|
|
887
|
+
|
|
888
|
+
**F14 carry-forward (universal):**
|
|
889
|
+
> Hər production cell üçün **yalnız o beat-də iştirak edən** elementlər attach olunur. "Color continuity / atmosphere / context üçün optional" QADAĞAN — model elementləri sızdırır.
|
|
890
|
+
|
|
891
|
+
**Plain language carry-forward (CLAUDE.md sərt qaydası):**
|
|
892
|
+
> AZ mətin təsvirdə texniki kinematik termin QADAĞAN: MCU, ECU, rule of thirds, focal length numbers (50mm), aperture (f/2.8), color temperature (3200K), angle terms (3/4, eye-level). Plain AZ əvəzlər istifadə olunur. İngilis model prompt-da texniki termin qalır.
|
|
893
|
+
|
|
894
|
+
### Knowledge updates
|
|
895
|
+
- [x] CLAUDE.md "Plain-language təsvir qaydası" əlavə olundu
|
|
896
|
+
- [x] CLAUDE.md "Reference attachment format qaydası" əlavə olundu
|
|
897
|
+
- [x] CLAUDE.md "Cell beat-strict isolation qaydası" əlavə olundu
|
|
898
|
+
- [x] storyboard-builder SKILL.md plain language enforcement
|
|
899
|
+
- [x] image-prompt-engineer SKILL.md identity lock format + cell isolation
|
|
900
|
+
- [x] F13 (reference tag syntax) + F14 (beat leakage) test-learnings.md-ə qeyd olundu
|
|
901
|
+
|
|
902
|
+
### Statistical update
|
|
903
|
+
|
|
904
|
+
| Test | Iter | Result | Knowledge |
|
|
905
|
+
|---|---|---|---|
|
|
906
|
+
| P-01 → P-09 + Pilləkən | various | 10 tests | F8-F12, Cultural research, FULL PROMPT, Numeric Ratio, Project Style Lock |
|
|
907
|
+
| **Reels Cell 2** | **v1 FAILED** | **3 critical issues** | **F13 (identity tag syntax), F14 (beat leakage), Plain language rule** |
|
|
908
|
+
|
|
909
|
+
### Bu failure-in dəyəri
|
|
910
|
+
Cell 2 v1 production failure göstərdi:
|
|
911
|
+
1. ✅ Reference attach syntax fundamental həll problem-dir — fayl adı tag-i identity bind yaratmır
|
|
912
|
+
2. ✅ Optional reference attach beat boundary pozur — minimal attach prinsipi sərt olmalıdır
|
|
913
|
+
3. ✅ Plain language description bağımsız sərt qaydadır — istifadəçi anlama yükü vacibdir
|
|
914
|
+
4. ✅ Multi-attach scenes (3+ refs) identity-lock dili olmadan generic output verir
|
|
915
|
+
|
|
916
|
+
3 yeni qayda CLAUDE.md-ə əlavə olundu, sistem öyrəndi.
|
|
917
|
+
|
|
918
|
+
---
|
|
919
|
+
|
|
920
|
+
*Versiya: 1.9 | Test sayı: 10 + Reels Cell 2 v1 | Yeni patterns: F13, F14, Plain language rule | Son yenilənmə: 2026-05-15*
|
|
921
|
+
|
|
922
|
+
---
|
|
923
|
+
|
|
924
|
+
## Entry: Reels Cell 2 v2 (Edit-mode mental model kəşfi)
|
|
925
|
+
|
|
926
|
+
**Tarix:** 2026-05-15
|
|
927
|
+
**Test növü:** Real production
|
|
928
|
+
**Iteration:** v2 — **HƏLƏ DƏ FAILED** (identity/clothing drift, balcony elements fərqli)
|
|
929
|
+
|
|
930
|
+
### Failure description
|
|
931
|
+
|
|
932
|
+
Cell 2 v2-də mən "FULL PROMPT" və "PRIMARY IDENTITY LOCK" language tətbiq etmişdim, lakin yenə də output:
|
|
933
|
+
- Ayşənin **clothing** fərqli idi (sage-green floral dress, ref-də pink dress with Peter Pan collar olduğu halda)
|
|
934
|
+
- Balcony **railing geometry** fərqli idi (vertical bars sayı və geometry ref-dən fərqli)
|
|
935
|
+
- Generic Pixar girl-ə yaxınlaşır identity baxımından
|
|
936
|
+
|
|
937
|
+
### Root cause (fundamental architecture problem)
|
|
938
|
+
|
|
939
|
+
**Mən generate-mode prompt yazırdım** (1000 söz uzun deskriptiv) refs attached olsa da. Bu sistemli olaraq səhvdir.
|
|
940
|
+
|
|
941
|
+
### Industry research kəşfləri (5 model docs)
|
|
942
|
+
|
|
943
|
+
**OpenAI Cookbook (gpt-image-1.5 prompting guide):**
|
|
944
|
+
- `images.edit()` endpoint (NOT generate) refs üçün
|
|
945
|
+
- `input_fidelity="high"` parametr face preservation üçün KRİTİK
|
|
946
|
+
- "**Only the first one you provide is preserved with extra richness**"
|
|
947
|
+
- Preserve list pattern: "Do not change X. Preserve Y. Replace only Z."
|
|
948
|
+
|
|
949
|
+
**Black Forest Labs (Flux Kontext docs):**
|
|
950
|
+
- Position labels ("image 1", "image 2") QADAĞAN — model bunları görmür
|
|
951
|
+
- Visual content description ("the woman with curly dark hair") istifadə et
|
|
952
|
+
- Horizontal composite — left/center/right spatial language
|
|
953
|
+
- 0.92 cosine similarity 6 successive edits (industry leader)
|
|
954
|
+
|
|
955
|
+
**Google DeepMind (Nano Banana Pro):**
|
|
956
|
+
- Karakter adlandırma məcburi — "Sarah", "John"
|
|
957
|
+
- Eyni token reuse hər generasiyada ("emerald eyes" hər dəfə)
|
|
958
|
+
- 14 image input, 5 person consistency
|
|
959
|
+
|
|
960
|
+
**Midjourney v7:**
|
|
961
|
+
- `--oref <URL> --ow 150` — omni weight 0-1000
|
|
962
|
+
- Single character ref; 2x GPU cost
|
|
963
|
+
|
|
964
|
+
**Ideogram v3:**
|
|
965
|
+
- Single `character_reference.image_url`
|
|
966
|
+
- Text rendering kralı
|
|
967
|
+
|
|
968
|
+
### Universal Edit-mode pattern (yeni standart)
|
|
969
|
+
|
|
970
|
+
```
|
|
971
|
+
[ONE-SENTENCE scene action]
|
|
972
|
+
|
|
973
|
+
PRESERVE EXACTLY:
|
|
974
|
+
- [features from first ref]
|
|
975
|
+
- [features from second ref]
|
|
976
|
+
- [visual style]
|
|
977
|
+
|
|
978
|
+
REPLACE ONLY:
|
|
979
|
+
- [pose / expression / lighting / camera]
|
|
980
|
+
|
|
981
|
+
DO NOT ADD:
|
|
982
|
+
- [forbidden elements]
|
|
983
|
+
```
|
|
984
|
+
|
|
985
|
+
**Maksimum 200 söz.** Uzun prompt = identity drift.
|
|
986
|
+
|
|
987
|
+
### F15 — Generate-mode prompt with refs causes drift (yeni failure pattern)
|
|
988
|
+
|
|
989
|
+
**Trigger:** 500+ söz deskriptiv prompt + refs attached.
|
|
990
|
+
**Səbəb:** Long generative prompt refs-i override edir, model "from scratch" generation rejiminə düşür.
|
|
991
|
+
**Düzəliş:** Edit-mode rewrite — 100-200 söz + preserve list + first-image priority.
|
|
992
|
+
|
|
993
|
+
### F16 — Missing preserve list = arbitrary drift
|
|
994
|
+
|
|
995
|
+
**Trigger:** Refs attached, lakin promptda preserve list yox.
|
|
996
|
+
**Səbəb:** Model nəyin saxlanmasını bilmir, prompt-da yazılmamış elementləri "redesign" edir.
|
|
997
|
+
**Düzəliş:** Açıq preserve list — clothing, accessories, location elements, visual style.
|
|
998
|
+
|
|
999
|
+
### Knowledge updates
|
|
1000
|
+
- [x] Yeni knowledge fayl: `skills/image-prompt-engineer/knowledge/multi-reference-workflows.md` (industry-correct 5-model workflows)
|
|
1001
|
+
- [x] CLAUDE.md "Reference attachment format" yenidən yazıldı — Edit-mode mental model
|
|
1002
|
+
- [x] image-prompt-engineer SKILL.md Addım 5 yenidən yazıldı
|
|
1003
|
+
- [x] storyboard-builder SKILL.md Production Cells bölməsi yenidən yazıldı
|
|
1004
|
+
- [x] F15 + F16 patterns əlavə olundu
|
|
1005
|
+
- [x] TODO.md "Edit-mode mental model qaydası" task əlavə olundu
|
|
1006
|
+
|
|
1007
|
+
### Bu failure-in dəyəri (fundamental architecture)
|
|
1008
|
+
Cell 2 v2 göstərdi ki, identity preservation **language polishing problem-i deyil** — **architecture problem-idir**. Generate-mode mental model refs ilə işləmir. Edit-mode mental model + universal preserve list + first-image priority = industry standard.
|
|
1009
|
+
|
|
1010
|
+
İndi sistem hər model üçün doğru yanaşmanı bilir.
|
|
1011
|
+
|
|
1012
|
+
---
|
|
1013
|
+
|
|
1014
|
+
*Versiya: 2.0 | Test sayı: 10 + Reels Cell 2 v1 + Cell 2 v2 | Yeni: F15, F16, Edit-mode mental model | Son yenilənmə: 2026-05-15*
|
|
1015
|
+
|
|
1016
|
+
---
|
|
1017
|
+
|
|
1018
|
+
## Cell 2 v3 → v4 (Reels girl-and-cat) — 2026-05-16
|
|
1019
|
+
|
|
1020
|
+
### Failure (v3): Composition divergence — full body shot when medium close-up requested
|
|
1021
|
+
|
|
1022
|
+
**Original prompt:** "Vertical 9:16, framed as a medium close-up showing her head, shoulders, and elbow with the balcony railing and a slice of golden-hour sky visible behind her. Camera at her seated eye-level."
|
|
1023
|
+
|
|
1024
|
+
**AI output (v3):** Tam bədən (full body) — başdan ayağa, feet/legs/lap görünür, kompozisiyanın ~50%-i sky və ətraf mühit. Camera slight low-angle (aşağıdan yuxarıya). Face upper-middle, strict upper-right deyil.
|
|
1025
|
+
|
|
1026
|
+
**Root cause:** AI image models (GPT-Image-2 daxil) **"medium close-up" terminini wider shot kimi interpret edir**. Statistical bias — training data-da "close-up" tag-i ilə işarələnmiş şəkillər real medium-close-up-dan **daha geniş** olur (web photography conventional). Prompt-da "head, shoulders, and elbow" konkretlik kifayət deyil — model "show the subject in context" default-a meyl edir.
|
|
1027
|
+
|
|
1028
|
+
### Fix strategy (v4 prompt):
|
|
1029
|
+
1. **CAPS aggressive override** — "TIGHT MEDIUM CLOSE-UP" + "chest-up ONLY"
|
|
1030
|
+
2. **Explicit exclusion** — "feet, legs, lap are OUTSIDE THE FRAME — they MUST NOT be visible"
|
|
1031
|
+
3. **Verify clause** — "If you see feet, legs, lap, or stool in the output, the framing is WRONG and must be re-rendered tighter"
|
|
1032
|
+
4. **Numeric precision** — "face occupies the UPPER-RIGHT THIRD of the frame"
|
|
1033
|
+
5. **Slim sky percentage** — "SLIM 15% slice of golden-hour sky" (rəqəm precision)
|
|
1034
|
+
6. **Camera height anti-default** — "exact horizontal eye-level (NOT below her looking up)"
|
|
1035
|
+
|
|
1036
|
+
### v4 result:
|
|
1037
|
+
**✅ Composition fixed** — tight medium close-up, no feet/legs/lap, chest-up framing perfect.
|
|
1038
|
+
**✅ No glowing windows** — distant buildings darkened correctly (v3 ⚠️ həll olundu).
|
|
1039
|
+
**⚠️ Camera height** — still slight low-angle (subtle), not strict horizontal.
|
|
1040
|
+
**⚠️ Face position** — upper-center, not strictly upper-right third.
|
|
1041
|
+
**⚠️ Gaze direction** — still slightly upward (NOT horizontal as instructed).
|
|
1042
|
+
**⚠️ Wilting geranium** — flowers still look fresh.
|
|
1043
|
+
**⚠️ Hair tuck behind ear** — both sides still loose.
|
|
1044
|
+
**⚠️ Inner brow** — slightly raised (concerned), not strict neutral.
|
|
1045
|
+
|
|
1046
|
+
### F17 — Framing instructions need aggressive numeric + visual override (yeni failure pattern)
|
|
1047
|
+
|
|
1048
|
+
**Trigger:** Prompt-da "medium close-up", "close-up", "tight frame" tipli kinematik termin istifadə olunur, lakin AI wider shot rendere edir.
|
|
1049
|
+
|
|
1050
|
+
**Səbəb:** Web photography training data-da "close-up" tag-i wider shots-la qarışıq markalandırılır. Pure verbal direction zəifdir.
|
|
1051
|
+
|
|
1052
|
+
**Düzəliş:**
|
|
1053
|
+
- CAPS instruction ("TIGHT", "MUST NOT", "ONLY")
|
|
1054
|
+
- Numeric specificity ("chest-up only", "15% slice", "upper-right THIRD")
|
|
1055
|
+
- Anti-default exclusion list ("feet MUST NOT appear, legs MUST NOT appear, lap MUST NOT appear")
|
|
1056
|
+
- Verify clause ("if X visible, framing is WRONG, re-render")
|
|
1057
|
+
|
|
1058
|
+
**Carry-forward rule:** Hər composition instruction-da minimum 3 element olmalıdır: (1) pozitiv tələb (TIGHT MEDIUM CLOSE-UP), (2) negativ exclusion list (NO feet, NO legs), (3) verify clause (if visible, regenerate).
|
|
1059
|
+
|
|
1060
|
+
### F18 — Gaze direction default: AI biases upward when "distant" specified
|
|
1061
|
+
|
|
1062
|
+
**Trigger:** Prompt-da "looking into distance", "distant unfocused point", "gazing past camera" — istifadə olunur, AI gözləri **yuxarı-sağa** (skyward) yönəldir.
|
|
1063
|
+
|
|
1064
|
+
**Səbəb:** Training data-da "looking into distance" + "sad expression" kombinasiyası tez-tez **upward gaze** ilə işarələnir (cinema cliché — sad person looking at sky). Horizontal distant gaze (sub-tekstal observed sadness) daha az təkrar olunur.
|
|
1065
|
+
|
|
1066
|
+
**Müşahidə:** Cell 2 v3 + v4 hər ikisində gaze upward-right idi, prompt v4-də "NOT upward, NOT skyward, NOT at sky — eyeline is HORIZONTAL" əlavə edildi, lakin **hələ də biraz yuxarı** qaldı. **2 iterasiyada həll olunmadı → model intrinsic limitation flag.**
|
|
1067
|
+
|
|
1068
|
+
**Düzəliş (multi-strategy):**
|
|
1069
|
+
1. **Specific anchor point** — "gazing at a distant building rooftop on the horizon line, exactly at her seated eye-level height"
|
|
1070
|
+
2. **Anti-default repeat** — "Her eyes are at the SAME HEIGHT as the horizon line in the background. Pupils centered horizontally in eye sockets, NOT tilted up"
|
|
1071
|
+
3. **2-iteration model limitation flag** — accept slight upward bias as model intrinsic, scene-level fix lazımdırsa landmark anchor istifadə et
|
|
1072
|
+
|
|
1073
|
+
**Carry-forward rule:** "Distant gaze" istənildikdə **specific anchor object** + "same height as horizon" mentioned olmalıdır. AI-yə "where to look" göstər, "where NOT to look" deyil.
|
|
1074
|
+
|
|
1075
|
+
### F19 — Wilting/aging effects under-rendered (vs fresh defaults)
|
|
1076
|
+
|
|
1077
|
+
**Trigger:** Prompt-da "wilting geranium", "wilting flowers", "drooping leaves", "faded fabric", "rust spots", "aged paint" tələb olunur. AI **fresh/new versiya** çəkir.
|
|
1078
|
+
|
|
1079
|
+
**Səbəb:** Training data-da florals + flowers tez-tez **commercial photography** (fresh, vibrant) ilə işarələnir. "Wilting" tag-i nadirdir, AI fallback fresh-ə düşür.
|
|
1080
|
+
|
|
1081
|
+
**Müşahidə:** Cell 2 v3 + v4 hər ikisində geranium fresh idi, profil "wilting" tələb edirdi. 2-iteration → model intrinsic limitation flag.
|
|
1082
|
+
|
|
1083
|
+
**Düzəliş:**
|
|
1084
|
+
1. **Multi-attribute description** — "wilting geranium — petals drooping downward, edges browning, leaves yellowing, stem bending under its own weight, faded color compared to fresh flowers"
|
|
1085
|
+
2. **Negative anchor** — "NOT a fresh florist's geranium — this plant has been neglected for weeks"
|
|
1086
|
+
3. **Reference comparison** — "compare to a 3-week-old neglected plant, not a garden center display"
|
|
1087
|
+
|
|
1088
|
+
**Carry-forward rule:** Aging/decay effects (wilting, rust, faded, peeling) AI-da default-da zəifdir. Multi-attribute description + comparison anchor + negative anchor məcburi.
|
|
1089
|
+
|
|
1090
|
+
### F20 — Subtle anatomical positioning under-rendered (hair tuck behind ear)
|
|
1091
|
+
|
|
1092
|
+
**Trigger:** Profil tələb edir ki, saç "one side tucked behind ear" olsun. AI **hər iki tərəfdən sərbəst saç** çəkir.
|
|
1093
|
+
|
|
1094
|
+
**Müşahidə:** Cell 2 v3 + v4 hər ikisində saç sərbəst, qulaq arxasında deyil. 2-iteration → model intrinsic limitation flag.
|
|
1095
|
+
|
|
1096
|
+
**Səbəb:** Tucking behavior asymmetric anatomical detail-dir, AI-da bu tip subtle posing rare-dir. Default-da "loose flowing hair" + "Pixar character design".
|
|
1097
|
+
|
|
1098
|
+
**Düzəliş:**
|
|
1099
|
+
1. **Explicit visual description** — "Her LEFT ear is fully visible (lobe and outer rim exposed) because that side of her hair is GENTLY TUCKED BEHIND THE EAR. The other side of her hair (right side) falls free past her shoulder. ONE ear visible, ONE ear covered."
|
|
1100
|
+
2. **Verify clause** — "If both ears are covered by hair, the tucking is INCORRECT — re-render with left ear exposed and right side free"
|
|
1101
|
+
|
|
1102
|
+
**Carry-forward rule:** Subtle anatomical positioning details (hair tucked, sleeve rolled, eyebrow raised) require explicit visual description + asymmetry callout + verify clause. Profile-level mention kifayət deyil.
|
|
1103
|
+
|
|
1104
|
+
### Knowledge updates (Cell 2 v3 → v4 sessiyası)
|
|
1105
|
+
- [x] F17 (composition aggressive override) — pattern əlavə olundu, **v4-də test edildi və işlədi**
|
|
1106
|
+
- [x] F18 (gaze direction default upward) — 2-iteration model limitation flag
|
|
1107
|
+
- [x] F19 (wilting/aging defaults to fresh) — 2-iteration model limitation flag
|
|
1108
|
+
- [x] F20 (hair tuck under-rendered) — 2-iteration model limitation flag
|
|
1109
|
+
- [ ] CLAUDE.md "Test learning capture qaydası" gücləndirilməlidir: validator Mərhələ 4 (knowledge capture) MƏCBURI, bypass edilə bilməz
|
|
1110
|
+
- [ ] image-validator SKILL.md Mərhələ 4 məcburi addım kimi yenidən sıralanmalıdır
|
|
1111
|
+
|
|
1112
|
+
### Sistemik öyrənmə (mənim öz davranışıma dair)
|
|
1113
|
+
Bu sessiyada istifadəçinin sualı ("hər səhvdən öyrənirsən?") mənim **passiv knowledge capture problemimi** aşkar etdi. Cell 2 v3 validator report-ından sonra mən test-learnings.md-yə **avtomatik yazmadım** — istifadəçi xatırlatmasaydı, F17-F20 patterns itəcəkdi. Bu, sistemik **manual-trigger problemidir**: qayda CLAUDE.md-də var, lakin praktikada hər sessiyada **mənim diqqətimə** bağlıdır. CLAUDE.md güclənməlidir — validator workflow-da Mərhələ 4 (knowledge capture) **bypass edilə bilməyən** addım olmalıdır. v4 indi yazıldı, sistemik düzəliş hələ qarşıdadır.
|
|
1114
|
+
|
|
1115
|
+
---
|
|
1116
|
+
|
|
1117
|
+
*Versiya: 3.0 | Test sayı: 10 + Reels Cell 2 v1-v2-v3-v4 | Yeni: F17, F18, F19, F20 + sistemik öz-öz öyrənmə | Son yenilənmə: 2026-05-16*
|
|
1118
|
+
|
|
1119
|
+
---
|
|
1120
|
+
|
|
1121
|
+
## Reels 04-heyet establishing v1 — 2026-05-16
|
|
1122
|
+
|
|
1123
|
+
### Result
|
|
1124
|
+
**0 ❌ + 9 ⚠️ — PRODUCTION-READY** (location identity 10/13, beat isolation 8/8, physical realism 7/7, style baseline consistent)
|
|
1125
|
+
|
|
1126
|
+
### F17 təkrarlanması — promotion threshold çatdı
|
|
1127
|
+
|
|
1128
|
+
**Cell 2 v3** — full body shot when medium close-up specified
|
|
1129
|
+
**Cell 2 v4** — still slight low-angle when horizontal eye-level specified
|
|
1130
|
+
**04-heyet establishing** — camera positioned BELOW standing eye-level when "~165cm standing adult eye-level" specified
|
|
1131
|
+
|
|
1132
|
+
**3 iterasiya, eyni pattern** — AI image models default-da **low-angle composition** seç edirlər, "eye-level" instruction-u zəif tutur. Promotion meyarı (3+ təkrar) çatdı.
|
|
1133
|
+
|
|
1134
|
+
### F17 — PROMOTED to master rule
|
|
1135
|
+
|
|
1136
|
+
**Master rule (image-prompt-engineer/knowledge/physical-realism.md və ya prompt-templates.md):**
|
|
1137
|
+
|
|
1138
|
+
> **Camera height + composition framing must use multi-layer enforcement.** Single-line instructions like "medium close-up" or "standing eye-level" are insufficient — AI defaults to wider shots and lower angles.
|
|
1139
|
+
>
|
|
1140
|
+
> **Required layers** (all 3 must be present):
|
|
1141
|
+
> 1. **Pozitiv tələb** (CAPS) — "TIGHT MEDIUM CLOSE-UP", "STANDING ADULT EYE-LEVEL ~165cm"
|
|
1142
|
+
> 2. **Numeric anchor** — "chest-up only", "~165cm", "15% sky slice"
|
|
1143
|
+
> 3. **Negative exclusion + verify** — "feet MUST NOT appear", "if camera is below standing eye-level, framing is WRONG"
|
|
1144
|
+
|
|
1145
|
+
### F21 — Sky percentage in establishing shots biases upward
|
|
1146
|
+
|
|
1147
|
+
**Trigger:** Prompt-da "slim slice of sky", "upper-right corner sky", "lower third concrete" tələb olunur. AI **upper-third dominant sky** rendere edir.
|
|
1148
|
+
|
|
1149
|
+
**Səbəb:** Establishing shot training data-da sunset/golden hour kompozisiyaları **sky-dominant** olur (atmospheric appeal). AI default-da "more sky = more appeal" bias-ı saxlayır.
|
|
1150
|
+
|
|
1151
|
+
**Müşahidə:** 04-heyet establishing — sky upper third dominant, prompt "slim slice" tələb etdi.
|
|
1152
|
+
|
|
1153
|
+
**Düzəliş:**
|
|
1154
|
+
1. **Specific percentage + visual anchor** — "Sky occupies ONLY the TOP 15% of the frame, between the building roofs and the upper edge. The concrete floor occupies the BOTTOM 60% of the frame."
|
|
1155
|
+
2. **Negative anchor** — "Sky is NOT a dominant element — it is a slim peek between the architecture"
|
|
1156
|
+
3. **Verify clause** — "If sky occupies more than 20% of the frame, re-render with tighter framing on the architecture"
|
|
1157
|
+
|
|
1158
|
+
**Carry-forward rule:** Establishing shots prompt-da sky percentage spesifik göstərilməlidir, "slim slice" yetərli deyil.
|
|
1159
|
+
|
|
1160
|
+
### Knowledge updates (04-heyet sessiyası)
|
|
1161
|
+
- [x] F17 promoted to master rule (3-iteration threshold)
|
|
1162
|
+
- [x] F21 yeni pattern əlavə olundu (sky percentage bias)
|
|
1163
|
+
- [x] Validator Mərhələ 3 məcburi — bu dəfə **atlamadım** (sistemik fix işlədi!)
|
|
1164
|
+
- [ ] image-prompt-engineer/knowledge/physical-realism.md-də F17 promoted rule məcburi qeyd olunmalıdır
|
|
1165
|
+
- [ ] Location-designer SKILL.md-də "Camera height enforcement" alt-qayda olmalıdır
|
|
1166
|
+
|
|
1167
|
+
### Sistemik müşahidə (mənim davranışıma dair — 2-ci instance)
|
|
1168
|
+
Bu dəfə validator nəticəsindən sonra **dərhal** knowledge capture etdim, **istifadəçi xatırlatmadan**. Sistemik fix (CLAUDE.md Mərhələ 3 məcburi) işləyir. F17 promotion 3-cü iterasiyada görünür çünki sistem indi hər iterasiyanı saxlayır.
|
|
1169
|
+
|
|
1170
|
+
---
|
|
1171
|
+
|
|
1172
|
+
*Versiya: 3.1 | Test sayı: 10 + Reels Cell 2 v1-v4 + 04-heyet v1 | Yeni: F17 PROMOTED, F21 | Sistemik fix işlədi — knowledge capture artıq atlanmır | Son yenilənmə: 2026-05-16*
|
|
1173
|
+
|
|
1174
|
+
---
|
|
1175
|
+
|
|
1176
|
+
## Reels Cell 3 v1 (Pişik həyətdə alert) — 2026-05-16
|
|
1177
|
+
|
|
1178
|
+
### Result
|
|
1179
|
+
**1 ❌ KRİTİK + 6 ⚠️** — orange-ginger calico patches DROPPED, cat now appears pure grey-white tabby.
|
|
1180
|
+
|
|
1181
|
+
### Wins
|
|
1182
|
+
- ✅ **F17 master rule sübut olundu** — camera CAT EYE-LEVEL doğru render edildi (3-iteration promotion-dan sonra ilk uğur)
|
|
1183
|
+
- ✅ Beat isolation 8/8 (heç bir növbəti cell elementi sızdırılmadı)
|
|
1184
|
+
- ✅ Physical realism 7/7
|
|
1185
|
+
- ✅ Location identity birə-bir match
|
|
1186
|
+
|
|
1187
|
+
### F22 — Subordinate color patches dropped in edit-mode (yeni failure pattern)
|
|
1188
|
+
|
|
1189
|
+
**Trigger:** Multi-color coat/pattern karakterləri (calico cat, tabby, dappled animals, multi-color clothing prints) edit-mode-da işləndikdə AI **dominant base color-u** saxlayır, **kiçik subordinate color patches-i** drop edir.
|
|
1190
|
+
|
|
1191
|
+
**Müşahidə (Cell 3 v1):**
|
|
1192
|
+
- Cat-ref-də: grey body + white chest/paws + **orange-ginger patches on head/back/tail base** (3 fərqli rəng zonası)
|
|
1193
|
+
- Generated image-də: yalnız grey + white (orange patches **TAMAMILƏ YOXDUR**)
|
|
1194
|
+
- Profile-də "orange patches = PRIMARY visual identifier" deyilmişdi
|
|
1195
|
+
|
|
1196
|
+
**Səbəb:** AI edit-mode "color simplification" bias — model dominant 2 rəngə fokuslanır, 3-cü rəng "noise" kimi qəbul edilir. Training data-da "grey-white cat" tag-i çox yaygındır, "calico" daha az; statistical bias dominant pattern-ə düşür.
|
|
1197
|
+
|
|
1198
|
+
**Düzəliş strategy (Cell 3 v2 üçün — F-N protect):**
|
|
1199
|
+
|
|
1200
|
+
1. **Repetition + emphasis** — orange patches 5+ dəfə təkrarlanır, hər dəfə fərqli kontekst:
|
|
1201
|
+
- "ORANGE-GINGER patches on head"
|
|
1202
|
+
- "ORANGE patches on shoulders"
|
|
1203
|
+
- "ORANGE patches across back"
|
|
1204
|
+
- "ORANGE patch at tail base"
|
|
1205
|
+
- "PRIMARY visual identifier — orange-ginger CALICO patches"
|
|
1206
|
+
|
|
1207
|
+
2. **Spatial anchor** — hər patch dəqiq yerdə:
|
|
1208
|
+
- "Above her left eye and forehead: a distinct orange-ginger patch"
|
|
1209
|
+
- "On her back, between shoulder blades: orange-ginger calico patch"
|
|
1210
|
+
- "At base of tail: orange-ginger band"
|
|
1211
|
+
|
|
1212
|
+
3. **Negative anchor + verify** — "She is NOT a pure grey-white tabby — she is a CALICO with three distinct color zones: grey, white, and ORANGE-GINGER. If the output shows only grey and white without orange patches, the cat identity is WRONG, regenerate."
|
|
1213
|
+
|
|
1214
|
+
4. **Visual reference comparison** — "Like a typical Turkish street calico cat with three-color coat — NOT a grey tabby, NOT a Russian Blue"
|
|
1215
|
+
|
|
1216
|
+
**Carry-forward rule:** Multi-color characters (3+ color zones) require:
|
|
1217
|
+
- Each color zone explicitly named + spatial location
|
|
1218
|
+
- Repetition (5+) of subordinate colors
|
|
1219
|
+
- Negative anchor naming the "wrong" simplification (grey tabby, Russian Blue)
|
|
1220
|
+
- Verify clause
|
|
1221
|
+
|
|
1222
|
+
### F17 master rule — FIRST PASS SUCCESS
|
|
1223
|
+
|
|
1224
|
+
**Müşahidə:** Cell 3 v1-də camera CAT EYE-LEVEL doğru render edildi. Bu, F17 promoted master rule-un **ilk uğurlu testi**dir.
|
|
1225
|
+
|
|
1226
|
+
**İşləyən patterns:**
|
|
1227
|
+
- CAPS instruction: "TIGHT MEDIUM SHOT", "CAT EYE-LEVEL"
|
|
1228
|
+
- Numeric anchor: "approximately 35-40cm from the ground"
|
|
1229
|
+
- Negative exclusion: "NOT standing human eye-level, NOT bird's-eye, NOT extreme low-angle"
|
|
1230
|
+
- Verify clause: "If the camera is at standing human eye-level instead of cat eye-level, regenerate"
|
|
1231
|
+
|
|
1232
|
+
**Validation:** Cell 2 v3 → v4 → 04-heyet (3 ⚠️) → **Cell 3 v1 (✅)** — F17 multi-layer enforcement working.
|
|
1233
|
+
|
|
1234
|
+
### Knowledge updates (Cell 3 v1 sessiyası)
|
|
1235
|
+
- [x] F17 master rule validated — first pass success after promotion
|
|
1236
|
+
- [x] F22 yeni pattern (subordinate color patches dropped)
|
|
1237
|
+
- [x] F21 sky percentage threshold (~20%) — borderline ama acceptable
|
|
1238
|
+
- [ ] Cell 3 v2 prompt must apply F22 aggressive color preservation
|
|
1239
|
+
|
|
1240
|
+
### Sistemik müşahidə
|
|
1241
|
+
Bu dəfə də Mərhələ 3 atlanmadı — knowledge capture validator workflow-da məcburi addım kimi tətbiq olunur. 3-cü ardıcıl session, sistemik fix sabit işləyir.
|
|
1242
|
+
|
|
1243
|
+
---
|
|
1244
|
+
|
|
1245
|
+
*Versiya: 3.2 | Test sayı: 10 + Reels Cell 2 v1-v4 + 04-heyet + Cell 3 v1 | Yeni: F22 + F17 first success | Son yenilənmə: 2026-05-16*
|
|
1246
|
+
|
|
1247
|
+
---
|
|
1248
|
+
|
|
1249
|
+
## Reels Cell 3 v2 (Pişik həyətdə alert — calico fix) — 2026-05-16
|
|
1250
|
+
|
|
1251
|
+
### Result
|
|
1252
|
+
**0 ❌ + 4 ⚠️ — PRODUCTION-READY (excellent quality)**
|
|
1253
|
+
|
|
1254
|
+
### F22 — FIRST-SUCCESS VALIDATION ✅
|
|
1255
|
+
|
|
1256
|
+
**Trigger:** v1-də orange-ginger calico patches DROPPED (boz-ağ tabby kimi göründü). v2-də F22 aggressive strategy tətbiq olundu.
|
|
1257
|
+
|
|
1258
|
+
**v2 strategy işlədi:**
|
|
1259
|
+
- 6+ təkrar "ORANGE-GINGER patches"
|
|
1260
|
+
- 3 spatial anchor (head/shoulders/tail base)
|
|
1261
|
+
- Negative anchor (NOT pure grey-white tabby, NOT Russian Blue)
|
|
1262
|
+
- Visual comparison ("typical Turkish or Mediterranean street calico")
|
|
1263
|
+
- Verify clause ("if no orange patches visible, regenerate")
|
|
1264
|
+
- Light interaction ("orange patches catch warm light, glow slightly amber")
|
|
1265
|
+
|
|
1266
|
+
**Generated image v2:**
|
|
1267
|
+
- ✅ Orange-ginger on head/temple area
|
|
1268
|
+
- ✅ Orange-ginger on left shoulder (visible)
|
|
1269
|
+
- ✅ Orange-ginger across back (visible band)
|
|
1270
|
+
- ✅ Orange tinge inside left ear
|
|
1271
|
+
- ✅ Three distinct color zones — grey, white, orange — clearly visible
|
|
1272
|
+
|
|
1273
|
+
**Conclusion:** F22 aggressive multi-layer color preservation strategy WORKS. Should be applied to all future multi-color subject prompts (clothing prints, mixed-fur animals, multi-color objects).
|
|
1274
|
+
|
|
1275
|
+
### F17 — SECOND CONSECUTIVE SUCCESS
|
|
1276
|
+
|
|
1277
|
+
Cell 3 v1 ✅ + Cell 3 v2 ✅ — camera cat-eye-level (~35-40cm) stable across two iterations. F17 master rule artıq **reliable production tool**-dur.
|
|
1278
|
+
|
|
1279
|
+
### F21 sky percentage — partial improvement
|
|
1280
|
+
Sky still ~22% (borderline). Apartment building upper portion dominant. Not blocking, but pattern persists. Next iteration: explicit "building roofline at ~80% frame height" anchor.
|
|
1281
|
+
|
|
1282
|
+
### Color identity recurring issues — Eye color
|
|
1283
|
+
**Pattern observed (Cell 3 v1 + v2):** Eyes rendered as yellow/yellow-green (v1) or amber-yellow (v2) — never pure "amber-gold" as profile requires.
|
|
1284
|
+
|
|
1285
|
+
**Hypothesis:** AI bias — "cat eyes" training data dominated by yellow/green tones; "amber-gold" rare specific tag. Similar to F22 pattern (color simplification toward dominant training distribution).
|
|
1286
|
+
|
|
1287
|
+
**Future fix (if needed):** Apply F22-style aggressive enumeration:
|
|
1288
|
+
- "Amber-gold eyes (warm orange-gold tone, like honey or whiskey — NOT yellow, NOT green, NOT lime)"
|
|
1289
|
+
- Visual comparison: "like a golden retriever's eye color"
|
|
1290
|
+
- Negative anchor + verify clause
|
|
1291
|
+
|
|
1292
|
+
Not promoted to F-N yet (single character, 2 iterations) — needs broader observation before pattern claim.
|
|
1293
|
+
|
|
1294
|
+
### Knowledge updates (Cell 3 v2 sessiyası)
|
|
1295
|
+
- [x] F22 first-success validated (production tool ready)
|
|
1296
|
+
- [x] F17 second-success (master rule stable across multiple iterations)
|
|
1297
|
+
- [ ] F21 still needs improvement (sky anchor enforcement)
|
|
1298
|
+
- [ ] Eye color pattern observed but not yet promoted
|
|
1299
|
+
|
|
1300
|
+
### Sistemik müşahidə
|
|
1301
|
+
4-cü ardıcıl sessiyada Mərhələ 3 (knowledge capture) avtomatik icra olunur. Sistem öz-özünə öyrənmə loop-u istikrarlı işləyir:
|
|
1302
|
+
- Cell 3 v1 → F22 yeni pattern aşkar olundu
|
|
1303
|
+
- Cell 3 v2 → F22 fix sübut olundu
|
|
1304
|
+
- Hər iterasiyada knowledge artır, gələcək promptlar güclənir
|
|
1305
|
+
|
|
1306
|
+
---
|
|
1307
|
+
|
|
1308
|
+
*Versiya: 3.3 | Test sayı: 10 + Reels Cell 2 v1-v4 + 04-heyet + Cell 3 v1-v2 | F22 production tool, F17 stable, eye color pattern observed | Son yenilənmə: 2026-05-16*
|
|
1309
|
+
|
|
1310
|
+
---
|
|
1311
|
+
|
|
1312
|
+
## Reels Cell 4 v1 (eye contact, 4 ref, vertical split) — 2026-05-16
|
|
1313
|
+
|
|
1314
|
+
### Result
|
|
1315
|
+
**❌ MAJOR FAIL — fundamental composition collapse**
|
|
1316
|
+
|
|
1317
|
+
### F23 — Vertical split / multi-level space compositions misinterpreted (yeni failure pattern — KRİTİK)
|
|
1318
|
+
|
|
1319
|
+
**Trigger:** Prompt-da "vertical split", "upper half = X / lower half = Y", "two levels of space", "third floor balcony + ground level courtyard simultaneously visible" tələb olunur.
|
|
1320
|
+
|
|
1321
|
+
**Müşahidə (Cell 4 v1):**
|
|
1322
|
+
- Prompt explicitly described vertical-split 9:16 frame: upper 50% = Ayşən on 3rd floor balcony looking down, lower 50% = cat in courtyard looking up, implied vertical eyeline
|
|
1323
|
+
- Image rendered: Ayşən balcony in upper-LEFT corner, courtyard occupying most of frame as single ground-level scene, cat at bottom-right
|
|
1324
|
+
- Balcony appears as if on ground level adjacent to courtyard, NOT 3 floors above
|
|
1325
|
+
- No vertical eyeline — Ayşən gazes upper-right, cat gazes upper-left (parallel, not converging)
|
|
1326
|
+
- Composition reduced to "two adjacent elements in single scene" instead of "two separated spaces in stacked composition"
|
|
1327
|
+
|
|
1328
|
+
**Root cause:** Image-to-image models training data dominated by **single-scene single-perspective** compositions. Multi-level isometric/cross-section compositions are RARE (mostly architectural diagrams, comic book panels). Statistical bias overwhelmingly to single-perspective renders. Prompt-level instruction insufficient to override training distribution.
|
|
1329
|
+
|
|
1330
|
+
**Why this is fundamentally different from previous F-patterns (F17/F22):**
|
|
1331
|
+
- F17 (composition framing) — fixable with multi-layer enforcement (CAPS + numeric + verify) ✅
|
|
1332
|
+
- F22 (color simplification) — fixable with explicit color enumeration + spatial anchors ✅
|
|
1333
|
+
- F23 (multi-level spatial composition) — **NOT fixable at prompt level** because the model cannot construct a coherent multi-perspective scene from a single image generation pass
|
|
1334
|
+
|
|
1335
|
+
**This is a model architecture limitation, not a prompt failure.**
|
|
1336
|
+
|
|
1337
|
+
### Fix strategy — split into shot-reverse-shot (film convention)
|
|
1338
|
+
|
|
1339
|
+
Standard film editing convention for eye contact across distance: **two separate shots cut together**, not one composite frame.
|
|
1340
|
+
|
|
1341
|
+
**Cell 4 restructure:**
|
|
1342
|
+
- **Cell 4a** — Ayşən's downward gaze (close-up, Ayşən primary, courtyard/cat softly visible below frame or implied)
|
|
1343
|
+
- **Cell 4b** — Cat's upward gaze (close-up, cat primary, balcony out of frame or implied above)
|
|
1344
|
+
|
|
1345
|
+
Eye contact established by **edit/cut**, not by single composite image.
|
|
1346
|
+
|
|
1347
|
+
In video assembly (Kling 3.0): the cut between 4a and 4b creates the eye contact emotionally. Each individual shot is composable by AI within its single-perspective bias.
|
|
1348
|
+
|
|
1349
|
+
### Carry-forward rule (F23 — MASTER RULE candidate)
|
|
1350
|
+
|
|
1351
|
+
**Never request multi-level spatial composition in a single image prompt.** When script requires eye contact / interaction between subjects in vertically/horizontally separated spaces, decompose into:
|
|
1352
|
+
1. **Subject A shot** — close-up of subject A, looking toward subject B's implied direction (subject B may be partially visible or out of frame)
|
|
1353
|
+
2. **Subject B shot** — close-up of subject B, looking toward subject A's implied direction
|
|
1354
|
+
3. **(Optional) Wide establishing** — both subjects in single scene at SAME spatial level (if narratively possible)
|
|
1355
|
+
|
|
1356
|
+
This applies to:
|
|
1357
|
+
- Eye contact across floors (balcony ↔ courtyard)
|
|
1358
|
+
- Eye contact across rooms (window ↔ outside)
|
|
1359
|
+
- Eye contact across distance (rooftop ↔ street)
|
|
1360
|
+
- Any prompt requesting "upper half = X / lower half = Y" or similar split-frame composition
|
|
1361
|
+
|
|
1362
|
+
### Other failures in Cell 4 v1
|
|
1363
|
+
- F22 calico patches regression — cat appeared pure grey-white again. Hypothesis: when prompt focuses on complex composition (4 refs + vertical split), individual identity preservation weakens. F22 enforcement may need REPETITION at the END of prompt as well as in PRESERVE section.
|
|
1364
|
+
- Ayşən's downward gaze instruction failed — she gazes upward-right (same as Cell 2). F18 (gaze direction bias upward) recurring.
|
|
1365
|
+
|
|
1366
|
+
### Knowledge updates (Cell 4 v1 sessiyası)
|
|
1367
|
+
- [x] F23 yeni master-candidate pattern (multi-level spatial composition fundamental limitation)
|
|
1368
|
+
- [x] F22 regression observed in complex multi-ref scenes
|
|
1369
|
+
- [x] F18 recurring (gaze direction bias)
|
|
1370
|
+
- [ ] Cell 4 must be restructured as 4a + 4b (shot-reverse-shot)
|
|
1371
|
+
- [ ] CLAUDE.md / image-prompt-engineer / storyboard-builder skills should encode F23 as ARCHITECTURAL rule (decompose multi-level scenes before prompting)
|
|
1372
|
+
|
|
1373
|
+
### Sistemik müşahidə
|
|
1374
|
+
Bu, 5-ci ardıcıl session-da Mərhələ 3 atlanmadı. Lakin F23 patterns əvvəlki F-N-lərdən fərqlidir — **fundamental limitation**, prompt rewriting ilə həll edilməz. Sistem öyrəndi ki, bəzən doğru cavab "prompt-u dəyişdir" deyil, "**arxitekturanı dəyişdir**" (1 cell → 2 cell shot-reverse-shot).
|
|
1375
|
+
|
|
1376
|
+
---
|
|
1377
|
+
|
|
1378
|
+
*Versiya: 3.4 | Test sayı: 10 + Reels Cell 2 v1-v4 + 04-heyet + Cell 3 v1-v2 + Cell 4 v1 | F23 fundamental + F22 regression + F18 recurring | Son yenilənmə: 2026-05-16*
|
|
1379
|
+
|
|
1380
|
+
---
|
|
1381
|
+
|
|
1382
|
+
## SİSTEMİK PROBLEM #2: Sərt CLAUDE.md qaydaları prompt yazarkən atlanır — 2026-05-16
|
|
1383
|
+
|
|
1384
|
+
### Trigger
|
|
1385
|
+
Contact sheet 3x3 storyboard prompt yazdım, **iki sərt CLAUDE.md qaydasını pozdum**:
|
|
1386
|
+
1. **"Reference image workflow qaydası"** (sərt) — hər generation üçün ATTACH bölməsi + ref tag mapping məcburi. Mən "Faza 1 establishing baseline" deyə düşünüb refs-i atladım. Yanlış.
|
|
1387
|
+
2. **"Storyboard kadr mətin təsviri qaydası"** (sərt, 2026-05-15 göstərişi) — hər panel üçün AZ plain mətin təsvir məcburi. Mən yalnız İngiliscə prompt yazdım, AZ təsvir atladım.
|
|
1388
|
+
|
|
1389
|
+
İstifadəçi sualı: "biz danışmışıq məncə, niyə yoxsan?" — sərt qaydalar var, mən atlamışam.
|
|
1390
|
+
|
|
1391
|
+
### Root cause
|
|
1392
|
+
**Mərhələ 3 sistemik fix (knowledge capture)** validation workflow-da işləyir, lakin **prompt yazarkən** CLAUDE.md sərt qaydalarını check-list kimi yoxlamaq prosesi YOXDUR. Hər prompt yazılarkən mən "bu sessiyada nə öyrəndik" düşünürəm, lakin "**CLAUDE.md hansı sərt qaydaları tələb edir bu növ promptdan**" sistematik yoxlamıram.
|
|
1393
|
+
|
|
1394
|
+
### Comparison — iki sistemik problem
|
|
1395
|
+
| Problem | Trigger | Sistemik fix |
|
|
1396
|
+
|---|---|---|
|
|
1397
|
+
| Mərhələ 3 (knowledge capture) atlanır | Validator verdikt verir, knowledge atlanır | CLAUDE.md "Mərhələ 3 məcburi" qaydası + image-validator SKILL.md addım yenidən sıralandı |
|
|
1398
|
+
| Sərt CLAUDE.md qaydaları prompt yazarkən atlanır | Yeni prompt yazılır, sərt qaydalar check edilmir | **HƏLƏ FIX EDILMƏYIB** — sistematik check-list yoxdur |
|
|
1399
|
+
|
|
1400
|
+
### Pozulan sərt qaydalar (bu sessiyada)
|
|
1401
|
+
1. Reference image workflow — Contact sheet v1-də refs yox idi (v2-də 4 ref əlavə edildi)
|
|
1402
|
+
2. Storyboard kadr AZ mətin təsviri — Contact sheet v1-də 9 panel təsviri yox idi (v2-də əlavə edildi)
|
|
1403
|
+
|
|
1404
|
+
### Fix strategy (təklif olunur)
|
|
1405
|
+
|
|
1406
|
+
**Option A — Skill SKILL.md addımlarında məcburi check-list:**
|
|
1407
|
+
- `storyboard-builder` SKILL.md Addım N: "Prompt yazmazdan ƏVVƏL CLAUDE.md sərt qaydaları yoxla: Reference workflow + AZ mətin təsvir + Plain language + Beat isolation"
|
|
1408
|
+
- `image-prompt-engineer` SKILL.md Addım N: eyni check-list
|
|
1409
|
+
|
|
1410
|
+
**Option B — CLAUDE.md-də "Prompt yazma protokolu" qaydası:**
|
|
1411
|
+
- Hər prompt yazmazdan əvvəl 5-element check-list:
|
|
1412
|
+
1. ATTACH bölməsi (refs məcburi)
|
|
1413
|
+
2. AZ plain mətin təsvir (panel/cell üçün)
|
|
1414
|
+
3. Beat-strict isolation (hansı element OLMAYACAQ)
|
|
1415
|
+
4. Edit-mode struktur (PRESERVE/REPLACE/DO NOT ADD)
|
|
1416
|
+
5. Verify clauses (regenerate triggers)
|
|
1417
|
+
|
|
1418
|
+
**Option C — Prompt template fayllar:**
|
|
1419
|
+
- `tests/_templates/storyboard-contact-sheet.template.md` — boş template ATTACH bölməsi və AZ təsvir slot-lar ilə
|
|
1420
|
+
- `tests/_templates/production-cell.template.md` — boş template ATTACH + AZ + edit-mode struktur
|
|
1421
|
+
- Hər yeni prompt template-dən doldurulur, atlanma imkansız
|
|
1422
|
+
|
|
1423
|
+
### İstifadəçi tələb edir hansı fix
|
|
1424
|
+
"Bu sərt qaydadır biz danışmışıq" — yəni qayda artıq mövcuddur (CLAUDE.md-də), tətbiqdə atlanır. Fix sistemik tətbiq səviyyəsində olmalıdır, qayda yenidən yazılmamalı.
|
|
1425
|
+
|
|
1426
|
+
### Carry-forward rule
|
|
1427
|
+
**Hər prompt yazmazdan əvvəl (storyboard, production cell, video, audio) — CLAUDE.md sərt qaydaları check-list kimi yoxla:**
|
|
1428
|
+
- Reference attachment format (`@img1 = filename.png` qadağan, ATTACH bölməsi məcburi)
|
|
1429
|
+
- Storyboard kadr AZ mətin təsviri (sərt — hər panel üçün)
|
|
1430
|
+
- Plain-language AZ təsvirdə (texniki söz qadağan)
|
|
1431
|
+
- Cell beat-strict isolation
|
|
1432
|
+
- Physical realism
|
|
1433
|
+
- Edit-mode (refs varsa)
|
|
1434
|
+
- Prompt çatdırılma (chat-də inline code block)
|
|
1435
|
+
- FULL PROMPT (Fix mode-da fragment qadağan)
|
|
1436
|
+
|
|
1437
|
+
### Knowledge updates
|
|
1438
|
+
- [x] Contact sheet v2 düzəldildi (refs + AZ təsvir əlavə olundu)
|
|
1439
|
+
- [x] Sistemik problem #2 qeyd olundu (bu entry)
|
|
1440
|
+
- [ ] CLAUDE.md "Prompt yazma protokolu" qaydası əlavə olunmalıdır (5-element check-list)
|
|
1441
|
+
- [ ] `storyboard-builder`, `image-prompt-engineer`, `video-prompt-engineer` SKILL.md-lərdə prompt yazma Addımı əvvəlində CLAUDE.md check-list yoxlamağı məcburi qoy
|
|
1442
|
+
|
|
1443
|
+
### Bu sessiyada öyrəndiyim 2 sistemik problem (yekun)
|
|
1444
|
+
1. Mərhələ 3 knowledge capture passiv idi (DÜZƏLDİLDİ — CLAUDE.md güclənib, validator workflow yenidən sıralandı)
|
|
1445
|
+
2. Sərt CLAUDE.md qaydaları prompt yazarkən atlanır (HƏLƏ FIX EDILMƏYIB — Option B + C tövsiyə olunur)
|
|
1446
|
+
|
|
1447
|
+
---
|
|
1448
|
+
|
|
1449
|
+
*Versiya: 3.5 | Test sayı: 10 + Reels Cell 2 v1-v4 + 04-heyet + Cell 3 v1-v2 + Cell 4 v1 + Contact sheet v1 (fail) → v2 | Sistemik problem #2 aşkar olundu: sərt qaydalar prompt yazma mərhələsində check edilmir | Son yenilənmə: 2026-05-16*
|
|
1450
|
+
|
|
1451
|
+
---
|
|
1452
|
+
|
|
1453
|
+
## SİSTEMİK PROBLEM #3: Multi-panel single-pass narrative continuity failure — 2026-05-16
|
|
1454
|
+
|
|
1455
|
+
### Trigger
|
|
1456
|
+
Contact sheet v2 generate olundu — 9 panel ssenari-uyğunsuz alındı:
|
|
1457
|
+
- Panel 4 (Beat 3a Ayşən downward) = Panel 2-nin demək olar kopyası, head turn yox
|
|
1458
|
+
- Panel 5 (Beat 3b Pişik upward) = Panel 3 ilə eyni, head tilt fərqli deyil
|
|
1459
|
+
- Panel 6 (Beat 4a Çörək alır) = Ayşən **dəftərə yazır** (çörək yox)
|
|
1460
|
+
- Panel 7 (Beat 4b Çörək atılır + cat leaps) = heç bir action freeze yox, statik
|
|
1461
|
+
- Panel 8 (Beat 5 Pilləkəndən qaçır) = qaçmır, dayanır + əlində çanta (ssenarıdan kənar)
|
|
1462
|
+
- Panel 9 (Beat 6 Qucaqda) = qucaqda yox, çömbəlir + pişik qarşıda
|
|
1463
|
+
|
|
1464
|
+
**6/9 panel ssenari-image disconnect (67% failure rate).**
|
|
1465
|
+
|
|
1466
|
+
### İstifadəçi tələbi
|
|
1467
|
+
"Storyboard ssenariyə görə ardıcıl kadrlar olmalıdır kadrlar bir birinin davamı olmalıdır dəxlisiz kadrlar olmamaldırı bu çox ciddi problemdir. Lazımdırsa dünyada araşdır mütləq."
|
|
1468
|
+
|
|
1469
|
+
### Industry research aşkar etdi (2026-05-16)
|
|
1470
|
+
|
|
1471
|
+
**Source 1 — Topview AI storyboard pipeline (medium.com):** "4-5 frames is the reliable range in a single generation, after that you're managing drift rather than preventing it. For 6+ shots, break into multiple sessions with reference images anchoring each new batch."
|
|
1472
|
+
|
|
1473
|
+
**Source 2 — Nano Banana Pro storyboard guide (sider.ai, 2026):** "The most reliable way to maintain consistency is using reference images. When a frame drifts, attaching your approved image to the new prompt anchors the model more strongly than text alone."
|
|
1474
|
+
|
|
1475
|
+
**Source 3 — CANVAS framework (arXiv 2604.13452):** Academic paper proposes "memory-guided sequential generation to retrieve relevant visual anchors across shots" — multi-panel coherence solution.
|
|
1476
|
+
|
|
1477
|
+
**Source 4 — GPT Image 2 documentation (vicsee.com, atlascloud.ai):** "Every clip in a storyboard-to-animation sequence is generated against an image that fixes character, lighting, wardrobe, environment, and lens. This makes the model's job a constraint problem rather than a generation problem from scratch."
|
|
1478
|
+
|
|
1479
|
+
### F24 — Multi-panel narrative continuity single-pass FAIL (MASTER LIMITATION)
|
|
1480
|
+
|
|
1481
|
+
**Promotion meyarı:** Sessiya kəşfi (Contact sheet v2) + industry validation 4 müstəqil mənbədən. Master pattern.
|
|
1482
|
+
|
|
1483
|
+
**Trigger:** Prompt-da 6+ panel single-pass tələb olunur (məs. 3×3 = 9, 4×2 = 8, 4×3 = 12). AI image models bunu **fundamental şəkildə bacarmır** — hər panel isolated still kimi rendere olunur, narrative arc + action progression saxlanılmır.
|
|
1484
|
+
|
|
1485
|
+
**Root cause:** AI image models training data-da single-scene generation dominantdır. Multi-panel narrative coherence training-dən gəlmir — model 6+ panel-də drift management problem-i yaranır. Bu prompt rewriting ilə həll olunmur — fundamental architectural limitation.
|
|
1486
|
+
|
|
1487
|
+
**Industry-validated həll (master rule):**
|
|
1488
|
+
|
|
1489
|
+
1. **Avoid** single-pass 6+ panel generation
|
|
1490
|
+
2. **Sequential cell-by-cell** workflow standard:
|
|
1491
|
+
- Cell N generation: globals (character ref + location ref) + previous cell image as continuity anchor
|
|
1492
|
+
- Validator → if ✅ → use as anchor for Cell N+1
|
|
1493
|
+
- Per cell isolated, single-perspective scene
|
|
1494
|
+
3. **Optional contact sheet compose** post-production (manual montage)
|
|
1495
|
+
4. **Style baseline exploration** — max 4 panel single-pass (look-and-feel only, not production)
|
|
1496
|
+
|
|
1497
|
+
**Carry-forward rule (MASTER):**
|
|
1498
|
+
- Production storyboard **always sequential cell-by-cell**, never single-pass 6+ panel mega-prompt
|
|
1499
|
+
- Each new cell anchored to previous approved cell image as visual reference
|
|
1500
|
+
- Skip "contact sheet" stage for production (it was based on misconception)
|
|
1501
|
+
|
|
1502
|
+
### Knowledge updates
|
|
1503
|
+
- [x] F24 PROMOTED to master rule (industry validation + 1 production failure)
|
|
1504
|
+
- [x] CLAUDE.md "Contact sheet feedback loop qaydası" YENİDƏN YAZILDI (sequential strategy)
|
|
1505
|
+
- [x] Storyboard-plan.md document approach replaces single-pass contact sheet
|
|
1506
|
+
- [ ] storyboard-builder SKILL.md must be updated (Faza 1 = planning doc, Faza 2 = sequential cells)
|
|
1507
|
+
- [ ] image-prompt-engineer SKILL.md must include "previous cell image as continuity anchor" pattern
|
|
1508
|
+
|
|
1509
|
+
### Reels Cell 2-Cell 3-Cell 4 case study (səbəb-nəticə)
|
|
1510
|
+
- Cell 2 v4 + Cell 3 v2 + 04-heyet establishing — hər biri sequential cell-by-cell idi → SUCCESS
|
|
1511
|
+
- Cell 4 v1 (vertical split) — tried multi-element single-pass → F23 fail
|
|
1512
|
+
- Contact sheet v2 (9 panel single-pass) — tried multi-panel → F24 fail
|
|
1513
|
+
- **Pattern:** Sequential = ✅, single-pass multi-element = ❌. Industry confirms.
|
|
1514
|
+
|
|
1515
|
+
### Sistemik öyrənmə
|
|
1516
|
+
Bu 3-cü sistemik problemdir bu sessiyada:
|
|
1517
|
+
1. Mərhələ 3 knowledge capture passiv → DÜZƏLDİLDİ (commit 17f953a)
|
|
1518
|
+
2. Sərt CLAUDE.md qaydaları prompt yazarkən atlanır → DÜZƏLDİLDİ (commit 090a734)
|
|
1519
|
+
3. Storyboard single-pass mega-prompt yanlış strategy idi → İNDİ DÜZƏLDİLİR (CLAUDE.md "Contact sheet feedback loop" yenidən yazıldı + F24 master rule + storyboard-builder workflow yenilənir)
|
|
1520
|
+
|
|
1521
|
+
---
|
|
1522
|
+
|
|
1523
|
+
## 2026-05-17 — Camal v1 (Stress Test #1 — 6-view Character Contact Sheet)
|
|
1524
|
+
|
|
1525
|
+
### Test context
|
|
1526
|
+
**Test type:** Multi-failure-pattern stress test — bir obrazda F1+F4+F6+F22 hamısını yoxlamaq
|
|
1527
|
+
**Model:** GPT-Image-2 (generate-mode, refs yox — initial baseline)
|
|
1528
|
+
**Layout:** 2+4 panel single image (full body front + back, plus 4 close-up angles)
|
|
1529
|
+
|
|
1530
|
+
### Failure summary (4 ❌, 3 ⚠️, 17 ✅)
|
|
1531
|
+
|
|
1532
|
+
**❌ F4 (Material distinction) — Mismatched elbow patches:**
|
|
1533
|
+
- **Prompt parça:** "LEFT ELBOW patch: SMOOTH BROWN LEATHER... RIGHT ELBOW patch: WOOL FELT... they MUST be different materials"
|
|
1534
|
+
- **Beklənilən nəticə:** Sol dirsəkdə dəri (parlaq), sağ dirsəkdə yun (mat) — açıq fərqlənən material
|
|
1535
|
+
- **Faktiki nəticə:** Hər iki dirsəkdə eyni material — leather-görünüşlü, mismatched DEYIL
|
|
1536
|
+
- **❌ Pozulma kateqoriyası:** prompt-consistency F4 (material ignored — bilinən pattern, recurring)
|
|
1537
|
+
|
|
1538
|
+
**❌ F4 + F1 (Asymmetric material) — Mismatched boot soles:**
|
|
1539
|
+
- **Prompt parça:** "LEFT BOOT sole: BLACK rubber... RIGHT BOOT sole: BROWN LEATHER... clearly different colors"
|
|
1540
|
+
- **Beklənilən nəticə:** Sol qara, sağ qəhvəyi — açıq fərq
|
|
1541
|
+
- **Faktiki nəticə:** Hər ikisi eyni qaramtıl, mismatched DEYIL
|
|
1542
|
+
- **❌ Pozulma kateqoriyası:** F4 + F1 (asymmetry collapse — model symmetric pair preference)
|
|
1543
|
+
|
|
1544
|
+
**❌ F6 (Number ignored) — Saz string count:**
|
|
1545
|
+
- **Prompt parça:** "EXACTLY 7 strings... Do NOT render 8 strings (modern saz default), Do NOT render 6 strings (guitar default)"
|
|
1546
|
+
- **Beklənilən nəticə:** 7 sim countable
|
|
1547
|
+
- **Faktiki nəticə:** 5-6 sim görünür (guitar default)
|
|
1548
|
+
- **❌ Pozulma kateqoriyası:** F6 (number ignored — strong baseline pattern, recurring)
|
|
1549
|
+
|
|
1550
|
+
**❌ F6 (Number) — Tuning pegs:**
|
|
1551
|
+
- **Prompt parça:** "7 wooden tuning pegs visible, arranged in two rows"
|
|
1552
|
+
- **Faktiki nəticə:** Headstock-da 6 peg
|
|
1553
|
+
- **❌ Pozulma:** F6 (eyni problem strings ilə — tuning peg count saz strings ilə bağlıdır)
|
|
1554
|
+
|
|
1555
|
+
**⚠️ F22 (Multi-tone preservation) — Beard patch pattern softened:**
|
|
1556
|
+
- **Prompt parça:** "DISTINCT PATCH PATTERN — NOT a uniform mix: chin DARK 70%, sides LIGHT GREY 70%"
|
|
1557
|
+
- **Faktiki nəticə:** Var amma yumşaq, distinct olmayan — almost uniform mixed
|
|
1558
|
+
- **⚠️ Pattern:** Multi-tone preservation **softened** (F22 mid-strength) — F22 təkrarlanır, lakin daha az dramatic uğursuzluq
|
|
1559
|
+
|
|
1560
|
+
**⚠️ Verification failures:**
|
|
1561
|
+
- Eksik üst sağ kəsici diş: Ağız bağlı, görünmür — verify edilə bilmir (prompt-da "mouth slightly parted" instruction olmalıdır)
|
|
1562
|
+
- Age spots on temples: Faint, distinct yox — F1 sub-pattern (small detail softening)
|
|
1563
|
+
|
|
1564
|
+
### Root cause analizi
|
|
1565
|
+
|
|
1566
|
+
**F4 (Material distinction) recurring pattern:**
|
|
1567
|
+
- Model training data-sında "elbow patches" konsepti **uniform** kimi öyrənilib (jacket-də patch-lər **eyni** olur tipik)
|
|
1568
|
+
- "Mismatched" instruction abstract qalır — model konkret necə render edəcəyini bilmir
|
|
1569
|
+
- AI default: paired body parts (elbows, boots) → symmetric rendering
|
|
1570
|
+
- **Statistical bias:** "leather patch" sözü görüləndə hər iki dirsəyə tətbiq edilir
|
|
1571
|
+
|
|
1572
|
+
**F6 (Number) recurring — string instrument specific:**
|
|
1573
|
+
- "Saz" konsepti AI-da rare (yüksək Western bias)
|
|
1574
|
+
- Default fallback: guitar (6 strings) və ya lute (varies)
|
|
1575
|
+
- "EXACTLY 7" instruction abstract — model spesifik sayı vizual olaraq saxlamır
|
|
1576
|
+
- **Pattern:** "Exactly N" instruction strings/fingers/objects-də zəif tətbiq olunur
|
|
1577
|
+
|
|
1578
|
+
**F22 (Multi-tone) softening — beard:**
|
|
1579
|
+
- Training data-da beard-lər **uniform** salt-and-pepper kimi rendere olunur
|
|
1580
|
+
- "Patch pattern" instruction qismən anlaşılır (variation var), lakin **distinct sahə** kimi deyil
|
|
1581
|
+
- F22 yumşaq variant — Cell 3 calico kimi tam drop deyil, partial softening
|
|
1582
|
+
|
|
1583
|
+
### Fix tətbiqi (v2 üçün strategiya)
|
|
1584
|
+
|
|
1585
|
+
**Universal: spatial location anchoring:**
|
|
1586
|
+
- F4 üçün: "Look at the LEFT elbow specifically — describe the patch material. Look at the RIGHT elbow specifically — describe the patch material. They MUST be visually distinguishable." Spatial isolation tələb edilir.
|
|
1587
|
+
- F6 üçün: "Count the strings: 1, 2, 3, 4, 5, 6, 7 — exactly seven, not six (guitar count), not eight (modern saz)." Numbered enumeration prompt-da.
|
|
1588
|
+
- F22 üçün: Cell 3 v2 success pattern reuse — multiple repetitions (5+), spatial anchor, verify clause, "NOT uniform mix" negative anchor
|
|
1589
|
+
|
|
1590
|
+
**Negative anchors (Cell 3 v2 success-dən borc):**
|
|
1591
|
+
- "NOT a guitar (6 strings). NOT an oud (12 strings). NOT a 4-string folk instrument. EXACTLY 7 strings — saz tipic 7-string folk variant."
|
|
1592
|
+
- "NOT both elbow patches leather. NOT both wool. ONE leather, ONE wool — visibly different."
|
|
1593
|
+
- "NOT uniform mixed beard. Two distinct zones: dark chin + light sides."
|
|
1594
|
+
|
|
1595
|
+
**Verification clauses (regenerate triggers):**
|
|
1596
|
+
- "If the saz has 6 strings, regenerate with 7."
|
|
1597
|
+
- "If both elbow patches look the same material, regenerate with one leather and one wool."
|
|
1598
|
+
- "If both boot soles are the same color, regenerate with one black and one brown."
|
|
1599
|
+
- "If the beard is uniformly mixed, regenerate with distinct dark chin + light sides zones."
|
|
1600
|
+
|
|
1601
|
+
**Mouth open instruction:**
|
|
1602
|
+
- v1 prompt: "lips relaxed, slightly parted enough to show..." — "slightly" çox zəif
|
|
1603
|
+
- v2: "Mouth OPEN, top lip raised — the missing UPPER RIGHT INCISOR gap is CLEARLY VISIBLE in close-up panels (3, 5)"
|
|
1604
|
+
|
|
1605
|
+
### Carry-forward rule
|
|
1606
|
+
|
|
1607
|
+
**F4 (material distinction in symmetric body parts):**
|
|
1608
|
+
- Paired body parts (elbows, knees, shoulders, hands, feet) symmetric rendering AI default
|
|
1609
|
+
- Mismatched material override: spatial isolation + negative anchor + verify clause
|
|
1610
|
+
- Template: "Look at LEFT [part] specifically: [material A]. Look at RIGHT [part] specifically: [material B]. NOT same material. Verify: if both look identical, regenerate."
|
|
1611
|
+
|
|
1612
|
+
**F6 (specific instrument string count):**
|
|
1613
|
+
- Instrument string counts AI default to nearest common (guitar 6, lute 6-12 varies)
|
|
1614
|
+
- Specific count override: enumerated counting + negative anchor with alternatives + verify clause
|
|
1615
|
+
- Template: "Count the strings: 1, 2, ..., N — exactly N, not [N-1], not [N+1]. NOT a [common alternative]."
|
|
1616
|
+
|
|
1617
|
+
**F22 mid-strength (multi-tone softening):**
|
|
1618
|
+
- Cell 3 calico fix pattern applies: 5+ repetitions, spatial zones, negative anchor, verify clause
|
|
1619
|
+
- For mid-strength F22 (single body part variation), same template scales down
|
|
1620
|
+
|
|
1621
|
+
### Knowledge update tövsiyəsi
|
|
1622
|
+
- [x] test-learnings.md entry əlavə olundu (bu entry)
|
|
1623
|
+
- [ ] `prompt-consistency-checks.md` F4 təkrarı qeyd et — paired body parts symmetric rendering
|
|
1624
|
+
- [ ] `prompt-consistency-checks.md` F6 təkrar pattern — instrument string count default
|
|
1625
|
+
- [ ] F22 mid-strength variant açıq qeyd et — full drop vs softening fərqi
|
|
1626
|
+
|
|
1627
|
+
### Statistical update (test #11)
|
|
1628
|
+
F4, F6, F22 hamısı **bilinən recurring pattern**dir. Bu sayı dəyişir:
|
|
1629
|
+
- F4 (material): test sayı 4+, recurring
|
|
1630
|
+
- F6 (number): test sayı 4+, recurring
|
|
1631
|
+
- F22 (multi-tone): test sayı 3+ (Cell 3 calico, P-09 fresnel, indi Camal beard)
|
|
1632
|
+
|
|
1633
|
+
---
|
|
1634
|
+
|
|
1635
|
+
*Versiya: 3.7 | Test sayı: 11 (Camal v1 stress test əlavə olundu) | F22 promoted further confirmation | F4 + F6 recurring confirmation | Son yenilənmə: 2026-05-17*
|
|
1636
|
+
|
|
1637
|
+
---
|
|
1638
|
+
|
|
1639
|
+
## 2026-05-17 — Camal v2 (Fix Mode Test — 2-iteration model limitation kəşfi)
|
|
1640
|
+
|
|
1641
|
+
### Test context
|
|
1642
|
+
**v1 verdikt:** 4 ❌ + 3 ⚠️ (saz strings, elbow patches, boot soles, tuning pegs / beard pattern, missing tooth visibility, age spots)
|
|
1643
|
+
**v2 strategy:** Aqressiv override — F4 spatial isolation, F6 enumerated counting + negative alternatives, F22 Cell 3 calico success pattern reuse, MOUTH OPEN explicit
|
|
1644
|
+
|
|
1645
|
+
### v2 nəticəsi
|
|
1646
|
+
|
|
1647
|
+
**5 FIX SUCCESS (improvements):**
|
|
1648
|
+
- ✅ Mouth OPEN — instruction worked
|
|
1649
|
+
- ✅ Age spots distinct (especially in left profile close-up)
|
|
1650
|
+
- ✅ Left eye droop more pronounced
|
|
1651
|
+
- ✅ Hair visible (back close-up)
|
|
1652
|
+
- ⚠️→ Beard pattern partially improved (sides lighter, but chin-vs-sides not "VISIBLY DISTINCT")
|
|
1653
|
+
|
|
1654
|
+
**3 PERSISTENT FAILURES (same ❌ in v1 and v2):**
|
|
1655
|
+
- ❌ Saz strings: still 6 (despite enumerated counting 1,2,3,4,5,6,7 + 4 negative anchors + verify clause)
|
|
1656
|
+
- ❌ Mismatched elbow patches: still appear similar (despite spatial isolation + material specifics + verify clauses)
|
|
1657
|
+
- ❌ Mismatched boot soles: still similar dark rubber (despite "BLACK left, BROWN right" + verify clause)
|
|
1658
|
+
|
|
1659
|
+
**1 NEW FAILURE (v2-introduced):**
|
|
1660
|
+
- ❌ Cross-panel footwear inconsistency: Panel 1 = black dress shoes (low-cut), Panel 2 = hiking boots (mid-calf treaded) — DIFFERENT FOOTWEAR in front vs back of same subject
|
|
1661
|
+
|
|
1662
|
+
### Root cause analizi — 2-iteration model limitation flag
|
|
1663
|
+
|
|
1664
|
+
**3 persistent failures = GPT-Image-2 intrinsic limitations (not prompt failures):**
|
|
1665
|
+
|
|
1666
|
+
1. **F6 (specific instrument string count) — GPT-Image-2 LIMITATION:**
|
|
1667
|
+
- Aggressive prompt instructions (numbered enumeration, 4 negative alternatives, verify clauses) **failed to override** the model's training data bias toward 6-string defaults
|
|
1668
|
+
- Statistical bias: lute/saz instruments in training data → 6 strings dominant
|
|
1669
|
+
- Text-based numerical override has weak effect on visual rendering counts
|
|
1670
|
+
- **Conclusion:** Text-only prompting CANNOT reliably enforce specific instrument string counts in GPT-Image-2
|
|
1671
|
+
|
|
1672
|
+
2. **F4 (mismatched paired body parts material) — GPT-Image-2 LIMITATION:**
|
|
1673
|
+
- Spatial isolation + material specifics + 4 verify clauses **failed** to render mismatched elbow patches
|
|
1674
|
+
- AI conceptual symmetry: paired body parts (elbows, knees, hands, feet) → symmetric rendering bias
|
|
1675
|
+
- Text instruction "different materials" gets overridden by visual conceptual symmetry
|
|
1676
|
+
- **Conclusion:** Text-only prompting CANNOT reliably enforce mismatched materials on paired body parts in GPT-Image-2
|
|
1677
|
+
|
|
1678
|
+
3. **F4+F1 (mismatched paired item attributes) — GPT-Image-2 LIMITATION + cross-panel drift:**
|
|
1679
|
+
- Same symmetric default for boot soles
|
|
1680
|
+
- PLUS new emergent failure: cross-panel detail drift (different footwear between panels)
|
|
1681
|
+
- **Conclusion:** Single-image multi-panel renderings drift on fine asymmetric details between panels
|
|
1682
|
+
|
|
1683
|
+
### NEW PATTERN (F25 candidate) — Cross-panel detail drift in single-image multi-panel
|
|
1684
|
+
|
|
1685
|
+
**Observation:** When generating a multi-panel character contact sheet (6 panels in 1 image), CRITICAL fine details (footwear type, asymmetric materials, specific counts) **drift between panels** of the same subject.
|
|
1686
|
+
|
|
1687
|
+
- Panel 1 (front body): dress shoes
|
|
1688
|
+
- Panel 2 (back body): hiking boots
|
|
1689
|
+
|
|
1690
|
+
This is a NEW pattern, related to F24 (multi-panel narrative drift) but at **fine detail level** rather than narrative level.
|
|
1691
|
+
|
|
1692
|
+
**Promotion criteria:** 3+ tests showing this pattern. Currently 1 test (Camal v2). Watch for repeat in next multi-panel tests.
|
|
1693
|
+
|
|
1694
|
+
### Fix workflow — 3-cü iterasiya QADAĞAN (CLAUDE.md "2-iteration model limitation flag")
|
|
1695
|
+
|
|
1696
|
+
Per CLAUDE.md rule: "Eyni ❌ və ya ⚠️ 2 iterasiyada təkrar olunarsa, validator/image-prompt-engineer 3-cü iterasiya cəhdi əvəzinə model limitation flag qaldırır."
|
|
1697
|
+
|
|
1698
|
+
**Alternative model recommendations:**
|
|
1699
|
+
|
|
1700
|
+
| Problem | Recommended alternative | Why |
|
|
1701
|
+
|---|---|---|
|
|
1702
|
+
| Saz 7 strings | **Midjourney v7** (`--style stylized --v 7`) + manual reference image of real 7-string saz | MJ v7 has stronger instrument detail rendering, manual ref overrides training bias |
|
|
1703
|
+
| Mismatched elbow patches | **Flux Kontext** with image-to-image (real photo of mismatched jacket as reference) | Flux Kontext follows visual reference closer than GPT-Image-2 |
|
|
1704
|
+
| Mismatched boot soles | **Same approach (Flux Kontext + photo ref)** OR accept symmetric soles | If photo-realism critical, image-to-image; if not, accept |
|
|
1705
|
+
| Cross-panel footwear drift | **Generate panels separately**, manual compose post-production | Single-pass multi-panel cannot guarantee fine-detail consistency |
|
|
1706
|
+
|
|
1707
|
+
**Scene preconditioning recommendations:**
|
|
1708
|
+
|
|
1709
|
+
| Problem | Scene precondition |
|
|
1710
|
+
|---|---|
|
|
1711
|
+
| Saz strings | Choose 6-string instrument (matches AI default) OR accept ~6 strings rendering |
|
|
1712
|
+
| Mismatched elbow patches | Either both leather (accept symmetric) OR change to highly visible color difference (red patch vs green patch) instead of texture difference |
|
|
1713
|
+
| Mismatched boot soles | Either both same OR change to obviously visible asymmetry (one boot is hiking, other is dress shoe — but maintain consistency across panels via separate generation) |
|
|
1714
|
+
|
|
1715
|
+
### User decision required (CLAUDE.md rule)
|
|
1716
|
+
|
|
1717
|
+
After 2-iteration limitation flag, user decides:
|
|
1718
|
+
- (a) Accept ❌ as model limitation, document and move on
|
|
1719
|
+
- (b) Try alternative model (Midjourney v7 or Flux Kontext) — re-generate
|
|
1720
|
+
- (c) Change scene preconditioning (remove asymmetric details, simplify) — re-write prompt with adjusted spec
|
|
1721
|
+
- (d) Accept GPT-Image-2 baseline as good enough (5 fixes ✅, 3 persistent ❌)
|
|
1722
|
+
|
|
1723
|
+
### Carry-forward rule
|
|
1724
|
+
|
|
1725
|
+
**For character refs with asymmetric details:**
|
|
1726
|
+
- DEFAULT: avoid asymmetric paired body parts materials in prompts unless using reference image input
|
|
1727
|
+
- WARN: when prompt has "mismatched [paired items]", flag as potential model limitation BEFORE generation
|
|
1728
|
+
- TEMPLATE in showrunner: "Bu character profil-də asymmetric paired details var (məs. mismatched elbow patches). GPT-Image-2 bunu reliably render edə bilməyəcək. Tövsiyə: (a) reference image input istifadə et (Flux Kontext), (b) asymmetric detail-i remove et, (c) accept symmetric rendering"
|
|
1729
|
+
|
|
1730
|
+
**For specific instrument/object counts:**
|
|
1731
|
+
- DEFAULT: text-only count instructions are weak for instruments in GPT-Image-2
|
|
1732
|
+
- WARN: "EXACTLY N" instructions where N != common default → flag potential limitation
|
|
1733
|
+
- TEMPLATE: "Bu prompt-da exactly N strings/items var. GPT-Image-2 6-string default ilə render edə bilər. Yoxlanış: ref image at və ya alternative model istifadə et"
|
|
1734
|
+
|
|
1735
|
+
### Knowledge update tövsiyəsi
|
|
1736
|
+
- [x] test-learnings.md entry əlavə olundu (bu entry)
|
|
1737
|
+
- [ ] prompt-consistency-checks.md F4 + F6 üçün "GPT-Image-2 specific limitation" annotation əlavə et
|
|
1738
|
+
- [ ] F25 candidate pattern (cross-panel detail drift) — yeni test cycle-larda izlə
|
|
1739
|
+
- [ ] character-designer SKILL.md-də "asymmetric paired details" warning əlavə et (showrunner level)
|
|
1740
|
+
|
|
1741
|
+
---
|
|
1742
|
+
|
|
1743
|
+
*Versiya: 3.8 | Test sayı: 12 (Camal v2 — 2-iteration model limitation kəşfi) | F4 + F6 confirmed as GPT-Image-2 intrinsic limitations | F25 candidate pattern (cross-panel drift) izlənilir | Son yenilənmə: 2026-05-17*
|
|
1744
|
+
|
|
1745
|
+
---
|
|
1746
|
+
|
|
1747
|
+
## 2026-05-21 — Session log: 5 universal rules added from 28-may documentary
|
|
1748
|
+
|
|
1749
|
+
Bir session-da, 28-may documentary layihəsindən (k.1-k.4 cells və L-01/L-02/L-03A location refs üzərində iş) **5 universal qayda** toolkit-ə əlavə olundu. Hər biri konkret layihə pozulmasından öyrənildi və bütün gələcək layihələrə tətbiq olunur.
|
|
1750
|
+
|
|
1751
|
+
### Yeni qaydaların xülasəsi
|
|
1752
|
+
|
|
1753
|
+
1. **FACE DIVERSITY in multi-figure cells** (HARD RULE)
|
|
1754
|
+
- Triggered by: k.4 schoolgirls — 4 identical face template rendered across formation
|
|
1755
|
+
- Added to: `image-prompt-engineer/knowledge/character-anatomy.md` + `image-validator/knowledge/physical-realism-checks.md` § A5.8
|
|
1756
|
+
- Rule: AI default = single face template cloned; prompt must explicitly demand variation across 10 dimensions (face shape, eye placement, nose, mouth, complexion, hair, age, body, individual markers, expressions); 2+ similar faces in frame = ❌
|
|
1757
|
+
|
|
1758
|
+
2. **CELLS in FULL COLOR, never native sepia/B&W** (HARD RULE)
|
|
1759
|
+
- Triggered by: k.1-k.4 cell promptları native sepia silver-gelatin generation kimi yazılmışdı
|
|
1760
|
+
- Added to: `image-prompt-engineer/knowledge/color-grading.md` + `image-validator/knowledge/prompt-consistency-checks.md` F14
|
|
1761
|
+
- Rule: Sepia / B&W = post-production grade, NOT generation spec; script's "sepia archive photo" describes final timeline look, not generation. Generation in native sepia breaks reusability + anchor consistency + post-prod LUT chain
|
|
1762
|
+
|
|
1763
|
+
3. **HARD RULE #0 scope clarification — applies to L-XX refs, NOT cells** (clarification)
|
|
1764
|
+
- Triggered by: k.1 cell yazılarkən empty-location rule yanlış olaraq cell-ə tətbiq olundu (horseless cart, no driver, no pedestrians)
|
|
1765
|
+
- Added to: `location-designer/knowledge/location-design.md` + `image-prompt-engineer/knowledge/location-design.md` — new "HARD RULE #0 — what about CELLS?" subsection
|
|
1766
|
+
- Rule: empty-of-people applies ONLY to L-XX location anchor refs (reusability); cells MUST populate per script (horses, vehicles, pedestrians, crowds — every script element rendered)
|
|
1767
|
+
|
|
1768
|
+
4. **MANDATORY period + place costume research per kadr** (HARD RULE)
|
|
1769
|
+
- Triggered by: k.3 audience cell research-without yazılmışdı, "1900s" generic ilə kifayətləndi
|
|
1770
|
+
- Added to: `image-prompt-engineer/knowledge/character-anatomy.md` (new HARD RULE section) + `image-validator/knowledge/script-consistency-checks.md` (PP-1..PP-6 cədvəli)
|
|
1771
|
+
- Rule: Concrete year + place (not era/region) + ethnic composition + period dress per group/class/gender + event nuance + social rules; archive photos mandatory (not Wikipedia text alone); discoveries feed back into knowledge loop
|
|
1772
|
+
|
|
1773
|
+
5. **CINEMATIC LIVED MOMENTS, never posed group photos** (HARD RULE)
|
|
1774
|
+
- Triggered by: k.4 schoolgirls cell — stiff symmetric 3-row formation posed at camera; user feedback "Burda elə bil yığılıb şəkil çəkdirirlər. Bütün kadrlarda sanki biz o dövrün içindəymişik kimi kənardan baxırıq belə olmalıdır."
|
|
1775
|
+
- Added to: `image-prompt-engineer/knowledge/composition.md` (primary HARD RULE) + `image-prompt-engineer/knowledge/cinematography.md` (Director's pre-prompt mindset cross-reference)
|
|
1776
|
+
- Rule: Every cell = lived moment from fly-on-the-wall observer position; 5-step director mindset (script meaning → screenwriter intent → location analysis → observer position → mid-action moment); natural blocking (3/4 angles, unaware subjects, imperfection); language to avoid ("posed", "lined up", "all looking at camera"); language to use ("caught mid-X", "we observe from", "unaware of camera"); even "archive photograph" cells stay natural — sepia comes in post
|
|
1777
|
+
|
|
1778
|
+
### Cross-reference web created
|
|
1779
|
+
|
|
1780
|
+
These 5 rules **multiply together** — they only work when applied in concert:
|
|
1781
|
+
|
|
1782
|
+
```
|
|
1783
|
+
Period + place research → knows who was in the room + how they dressed
|
|
1784
|
+
↓
|
|
1785
|
+
FACE DIVERSITY → each of those people is a unique individual
|
|
1786
|
+
↓
|
|
1787
|
+
Lived moments → they cluster naturally per the era's social rules
|
|
1788
|
+
↓
|
|
1789
|
+
Cells in full colour → the full-chroma data carries to post grading
|
|
1790
|
+
↓
|
|
1791
|
+
HARD RULE #0 (cells) → the cell renders everyone the script names
|
|
1792
|
+
in (L-XX empty ref) + (character refs)
|
|
1793
|
+
```
|
|
1794
|
+
|
|
1795
|
+
### Files modified in this session
|
|
1796
|
+
|
|
1797
|
+
| File | Change |
|
|
1798
|
+
|---|---|
|
|
1799
|
+
| `image-prompt-engineer/knowledge/character-anatomy.md` | +2 HARD RULE sections (face diversity, period research) |
|
|
1800
|
+
| `image-prompt-engineer/knowledge/color-grading.md` | +1 HARD RULE section (cells in full colour) |
|
|
1801
|
+
| `image-prompt-engineer/knowledge/composition.md` | +1 HARD RULE section (cinematic lived moments) |
|
|
1802
|
+
| `image-prompt-engineer/knowledge/cinematography.md` | +1 director's-mindset cross-reference section |
|
|
1803
|
+
| `image-prompt-engineer/knowledge/location-design.md` | +1 HARD RULE #0 cell-scope clarification |
|
|
1804
|
+
| `location-designer/knowledge/location-design.md` | +1 HARD RULE #0 cell-scope clarification |
|
|
1805
|
+
| `image-validator/knowledge/physical-realism-checks.md` | +A5.8 (face diversity checkpoint), version bump to 1.8 |
|
|
1806
|
+
| `image-validator/knowledge/prompt-consistency-checks.md` | +F14 (cell sepia/B&W check), version bump to 1.2 |
|
|
1807
|
+
| `image-validator/knowledge/script-consistency-checks.md` | +Period+place accuracy section (PP-1..PP-6), version bump to 1.1 |
|
|
1808
|
+
| `image-validator/knowledge/test-learnings.md` | This session log entry |
|
|
1809
|
+
|
|
1810
|
+
### Meta-pattern
|
|
1811
|
+
|
|
1812
|
+
This session is the first time **5 universal rules** were added in a single project-session. The 28-may documentary's strictness (canonical history, mixed Canadian diaspora audience, period accuracy at the level of individual photographs) functions as an aggressive stress test for the toolkit. Future similar-stress projects expected to surface more universal rules.
|
|
1813
|
+
|
|
1814
|
+
---
|
|
1815
|
+
|
|
1816
|
+
*Versiya: 3.9 | Test sayı: 12 + 1 multi-rule session (28-may 2026-05-21) | 5 yeni universal rule kategorisi: face diversity, full-colour cells, HARD RULE #0 scope, period+place research, cinematic lived moments | Son yenilənmə: 2026-05-21*
|