@azerogluemin/ai-bootstrap 0.2.1

package/templates/skills/image-prompt-engineer/knowledge/models/gpt-image-2.md

@@ -0,0 +1,120 @@
+# GPT-Image-2 — Image Model Knowledge
+
+## ⚠️ HARD RULE — Prompt Length Budget
+
+GPT-Image-2 promptları üçün **char budget məcburidir**. Texniki tavan ≠ effektiv tavan.
+
+| Parametr | Dəyər |
+|---|---|
+| Technical ceiling (API qəbul edir) | **32,000 char** |
+| Praktik effektiv limit | **~3,000 char (~500 söz)** |
+| Tövsiyə hədəfi | **≤2,800 char** |
+| Aggressive cut tələbi | **>3,000 char** |
+| Critical regenerate threshold | **>4,000 char** |
+
+**Niyə:** Model **early-token bias** ilə işləyir — promptun ilk ~500-800 tokeni dominant olur, **~500 tokendən sonra** instruction-lar zəifləyir və ya tamamilə ignor olunur. 9,000 char promptda kritik detallar (refs, identity locks, anti-grit, lighting, anti-synthetic) sondadırsa, model onları **görməyəcək**. Daha pisi, **explicit "prompt too long" error** atır və generation fail olur.
+
+**Verification habit (məcburi):** Hər prompt yazandan sonra `wc -c` ilə yoxla:
+
+```bash
+cat prompt.txt | wc -c
+```
+
+Detection thresholds:
+- **≤2,500 char:** ✅ optimal
+- **2,500-3,000 char:** ⚠️ pass, lakin aggressive compression mümkündür
+- **3,000-4,000 char:** ⚠️ quality drop riski, kompress et
+- **>4,000 char:** ❌ aggressive cut LAZIMDIR
+- **>8,000 char:** ❌ "prompt too long" error riski, mütləq yenidən yaz
+
+**Compression strategy:** `knowledge/general/prompt-budget.md` faylına bax — universal cut/keep checklist + V1 (9168 char) → V2 (2713 char) before/after misalı.
+
+---
+
+## Provider
+OpenAI
+
+## Type
+Multimodal transformer (DALL-E successor)
+
+## Strengths
+
+- Detailed scene compositions (multiple subjects + environment)
+- Renders text/signs/captions reliably
+- Strong understanding of complex prompts
+- Good hands and fingers
+- Spatial relationship comprehension
+- Style adherence
+
+## Weaknesses
+
+- Sometimes "AI look" sneaks in (over-rendered)
+- Heavy stylistic bias toward "polished" aesthetic
+- Reference image support is limited compared to Nano Banana
+- Less photorealistic than Nano Banana for portraits
+- Slower generation (~10-15s)
+
+## Best Practices
+
+1. **Cinematic photograph framing** — start with this phrase
+2. Specify scene context first, then subject
+3. Use "shot in the style of [photographer/director]" sparingly
+4. Describe scene first (location, time, mood), then subject within it
+5. Avoid generic adjectives
+6. Specify aspect ratio clearly
+7. Lighting analysis is interpreted well
+8. Use natural prose, not bullet lists
+
+## Prompt Anatomy
+
+```
+Cinematic photograph: [Setting context — time, place, mood first]
++ [Subject in scene with action]
++ [Lighting analysis — key, fill, rim, color]
++ [Camera + lens specifics]
++ [Mood / style / color grading notes]
+```
+
+## Sample Prompt
+
+```
+Cinematic photograph: 1960s American diner interior at night,
+melancholic atmosphere, period accurate. A 35-year-old Azerbaijani
+man (athletic build, short dark hair, charcoal wool coat) sits by
+a rain-streaked window, looking outside with quiet melancholy.
+Warm tungsten ceiling lights at 3200K create soft rim on his
+shoulders, cool blue daylight spill from window behind him acts
+as fill. Shot on Sony A7 IV with 85mm at f/1.4. Melancholic mood,
+shallow depth of field, photorealistic, vintage film color grading
+referencing Kodak Portra 400, 9:16 vertical aspect ratio.
+```
+
+## Common Mistakes
+
+- Starting with subject only (works better with scene context first)
+- Over-stylization ("epic dramatic stunning")
+- Forgetting to mention "cinematic photograph"
+- Mixing director references that conflict
+
+## Aspect Ratio
+
+Specify clearly:
+- "9:16 vertical aspect ratio"
+- "16:9 widescreen"
+- "Square 1:1"
+- "Anamorphic 2.39:1 cinemascope"
+
+## Text Rendering
+
+GPT-Image-2 can render text in image:
+- "Neon sign reading 'OPEN' in red script"
+- "Newspaper headline visible: 'STORM WARNING'"
+- "Storefront with text 'PHARMACY' in 1950s typography"
+
+Use this strength when scene calls for it.
+
+## Pricing tier
+$$$ — higher than Nano Banana
+
+## Speed tier
+Medium — 10-15s

package/templates/skills/image-prompt-engineer/knowledge/models/ideogram-v3.md

@@ -0,0 +1,101 @@
+# Ideogram v3 — Image Model Knowledge
+
+## Provider
+Ideogram
+
+## Type
+Diffusion
+
+## Strengths
+
+- Best-in-class text rendering (most reliable)
+- Logo and brand identity work
+- Magic Prompt feature (auto-enhances prompts)
+- Multiple style presets
+- Aspect ratio flexibility
+- Reasonable cost
+
+## Weaknesses
+
+- Photorealism behind Nano Banana / Flux
+- Less cinematic look out-of-box
+- Limited control over photographic style
+- Less effective for portrait work
+
+## Best Practices
+
+1. Use for any prompt with text/typography
+2. Specify exact text in quotes
+3. Magic Prompt can enhance — but disable for precision
+4. Style presets limit but speed up consistency
+5. Aspect ratio explicit
+6. Brand colors via hex codes
+
+## Prompt Anatomy
+
+```
+[Subject + composition]
++ [Text content "in quotes"]
++ [Style preset reference]
++ [Color palette]
++ [Aspect ratio]
+```
+
+## Sample Prompts
+
+### Brand poster with text
+```
+Vintage-style movie poster for "MIDNIGHT KAFE", stylized illustration
+of a man's silhouette by a rainy window. Title "MIDNIGHT KAFE" in
+bold sans-serif at top. Tagline "where every story begins at the
+counter" in script font below. Color palette: deep burgundy #4A0E1E,
+warm cream #F5E6D3, gold accent #C9A961. Vertical 9:16 poster
+format. Photorealistic vintage style.
+```
+
+### Logo
+```
+Modern minimalist logo for "Atlas Bip" tech brand. Letter "A" formed
+from connected dots and lines representing data network. Color:
+electric blue #00B4D8 on white background. Geometric, scalable
+vector style. Square 1:1.
+```
+
+### Newspaper headline
+```
+Newspaper front page mockup. Headline reads "AI REVOLUTION REACHES
+LOCAL CREATORS" in bold serif. Subheading "How indie filmmakers
+are using AI tools" in lighter weight. Black ink on aged paper
+texture. Top-down view, 16:9 landscape.
+```
+
+## Text Rendering Tips
+
+- Put exact text in DOUBLE QUOTES
+- Specify font style: serif, sans-serif, script, display
+- Specify weight: bold, light, regular
+- Color of text explicitly
+
+## Style Presets
+
+- General
+- Realistic
+- Design
+- Anime
+- Render 3D
+
+Choose based on output needs.
+
+## Magic Prompt
+
+When ON, Ideogram enhances your prompt with style additions.
+When OFF, your prompt is followed literally.
+
+For precision work: Magic OFF
+For exploration: Magic ON
+
+## Pricing tier
+$$ — moderate
+
+## Speed tier
+Fast — 5-8s

package/templates/skills/image-prompt-engineer/knowledge/models/imagen-4.md

@@ -0,0 +1,73 @@
+# Imagen 4 — Image Model Knowledge
+
+## Provider
+Google
+
+## Type
+Diffusion
+
+## Strengths
+
+- Excellent text rendering (signs, captions, typography)
+- Photorealistic outputs
+- Strong prompt adherence (literal interpretation)
+- Clean, professional defaults
+- Good with technical/architectural subjects
+
+## Weaknesses
+
+- Less cinematic by default (more "stock photo" feel)
+- Conservative outputs (less risk-taking)
+- Limited stylization compared to Midjourney
+- Reference image support evolving
+
+## Best Practices
+
+1. Be explicit about cinematic intent ("cinematic photograph")
+2. Specify all elements clearly (Imagen interprets literally)
+3. Use camera + lens + aperture for depth
+4. Reference photographers for style ("in the style of Annie Leibovitz")
+5. Specify aspect ratio in prompt
+6. Lighting setup must be explicit
+7. Use for product/architectural shots strength
+
+## Prompt Anatomy
+
+```
+[Cinematic intent declaration]
++ [Subject + detailed physical description]
++ [Setting + environment]
++ [Lighting setup + color temp]
++ [Camera + lens + aperture]
++ [Style notes + aspect ratio]
+```
+
+## Sample Prompt
+
+```
+Cinematic photograph in 9:16 vertical aspect: a 35-year-old
+Azerbaijani man, athletic build, short dark wavy hair, green eyes,
+wearing a charcoal wool overcoat. Standing by a rain-streaked window
+in a 1960s American diner, looking outside with melancholic
+expression. Lighting: warm tungsten ceiling pendant at 3200K from
+upper-left creating rim on shoulders, cool blue spill from rain-
+streaked window at 5600K behind. Shot on Sony A7 IV with Cooke S4
+85mm lens at f/1.4. Photorealistic, cinematic color grading with
+warm shadows and cool highlights, shallow depth of field, slight
+film grain.
+```
+
+## Text Rendering Strength
+
+Imagen 4 excels at typography:
+- "Vintage neon sign reading 'OPEN' in red cursive script"
+- "Newspaper front page with headline 'STORM APPROACHES'"
+- "Restaurant menu board with daily specials in chalk handwriting"
+
+Use this for brand work.
+
+## Pricing tier
+$$ — moderate
+
+## Speed tier
+Medium — 5-10s

package/templates/skills/image-prompt-engineer/knowledge/models/midjourney-v7.md

@@ -0,0 +1,85 @@
+# Midjourney v7 — Image Model Knowledge
+
+## Provider
+Midjourney
+
+## Type
+Diffusion (proprietary, Discord/web-based)
+
+## Strengths
+
+- Aesthetic by default — produces "pretty" images effortlessly
+- Strong style consistency
+- Minimal prompting works (5-15 words can produce good output)
+- Active style community sharing
+- Character reference (`--cref`) and style reference (`--sref`) parameters
+- Stylize parameter (`--stylize`) for style intensity
+
+## Weaknesses
+
+- Discord/web UX (no robust API for automation)
+- Distinct "Midjourney look" — most outputs feel similar
+- Difficult to fully override default style
+- Can over-stylize (loses photorealism)
+- Less flexible for specific brand/photographic looks
+
+## Best Practices
+
+1. Short, evocative prompts (15-50 words) work better than long
+2. Use parameters at end: `--ar 9:16 --style raw --v 7 --stylize 250`
+3. Image references via URL: `[ref-url] cinematic portrait of ...`
+4. "Cinematic" is implied — don't over-specify
+5. Use `--style raw` for less stylization (more photo-realistic)
+6. `--cref [url]` for character consistency
+7. `--sref [url]` for style transfer
+8. Lower `--stylize` for realism, higher for art
+
+## Prompt Anatomy
+
+```
+[Reference URL if any] [Short subject description], [setting],
+[lighting/mood], [camera/lens reference] [parameters]
+```
+
+## Sample Prompt
+
+```
+A 35-year-old Azerbaijani man in charcoal wool coat by rain-streaked
+window of 1960s diner, melancholic, warm tungsten interior cool blue
+exterior, shot on ARRI with 85mm Cooke at f/1.4 --ar 9:16 --style raw
+--v 7 --stylize 100
+```
+
+## Parameters Reference
+
+| Param | Effect | Range |
+|-------|--------|-------|
+| `--ar` | Aspect ratio | 16:9, 9:16, 1:1, 21:9 |
+| `--v` | Version | 7 (current) |
+| `--style` | Style preset | raw, expressive, scenic, cute |
+| `--stylize` | Style intensity | 0-1000 (100 default) |
+| `--quality` | Quality | 0.25, 0.5, 1, 2 |
+| `--chaos` | Variation | 0-100 |
+| `--seed` | Reproducibility | integer |
+| `--cref` | Character reference | URL |
+| `--sref` | Style reference | URL |
+| `--no` | Negative prompt | "no [element]" |
+
+## Style Raw Note
+
+For photorealism, ALWAYS use `--style raw --stylize 100` or lower.
+Without these, Midjourney over-stylizes.
+
+## Common Mistakes
+
+- Long verbose prompts (Midjourney prefers concise)
+- High stylize for photorealistic intent
+- Forgetting `--style raw` for realism
+- No aspect ratio (defaults to 1:1)
+- Mixing too many style references
+
+## Pricing tier
+$$ — subscription based
+
+## Speed tier
+Fast — 30-60s for 4 variations

package/templates/skills/image-prompt-engineer/knowledge/models/nano-banana.md

@@ -0,0 +1,128 @@
+# Nano Banana — Image Model Knowledge
+
+## ⚠️ HARD RULE — Prompt Length Budget
+
+Nano Banana üçün **API limit və web UI limit fərqlidir** — və əksər istifadəçi web UI istifadə edir.
+
+| Parametr | Dəyər |
+|---|---|
+| API technical ceiling | **131,000 tokens (~500K char)** — Gemini 3.1 Flash Image |
+| API praktik effektiv | **~8,000 char** (early-token bias-dan sonra ignor) |
+| **Web UI hard limit** | **2,000 char** ← çox vaxt effektiv tavan |
+| Tövsiyə hədəfi (web UI istifadəçi) | **≤2,000 char** |
+| Tövsiyə hədəfi (API istifadəçi) | **≤2,800 char** |
+
+**KRİTİK:** Əksər istifadəçi prompt-ları **web UI vasitəsilə paste edir** — yəni 2,000 char effektiv tavandır. Bu limit aşılarsa, **silently truncate** olur (heç bir error mesajı yoxdur, lakin sondakı kritik instruction-lar — anti-synthetic clause, aspect ratio, identity lock — kəsilir).
+
+API istifadəsində belə, model **early-token bias** ilə işləyir — ilk ~500-800 token dominant olur, qalanı zəifləyir.
+
+**Verification habit (məcburi):** Hər prompt yazandan sonra `wc -c` ilə yoxla:
+
+```bash
+cat prompt.txt | wc -c
+```
+
+Detection thresholds (web UI istifadəçi üçün):
+- **≤1,800 char:** ✅ optimal
+- **1,800-2,000 char:** ⚠️ tam pasta olur, lakin marginda
+- **>2,000 char:** ❌ silent truncate baş verir — kritik instruction-lar son cümlədəndirsə itəcək
+- **>4,000 char:** ❌ API-də belə quality drop, kompress et
+
+**Compression strategy:** `knowledge/general/prompt-budget.md` faylına bax — universal cut/keep checklist + V1 (9168 char) → V2 (2713 char) before/after misalı.
+
+---
+
+## Provider
+Google
+
+## Type
+Diffusion + transformer hybrid
+
+## Strengths
+
+- Photorealistic portraits with natural skin texture
+- Strong character consistency with reference images (vision input)
+- Cinematic look out-of-box, minimal "AI feel"
+- Vision input — supports up to 5 reference images per prompt
+- Fast generation (~3-5s)
+- Excellent lighting handling
+- Good with anatomical anchors
+
+## Weaknesses
+
+- Cannot render text/signs/captions reliably
+- Hands and fingers occasionally distorted
+- Complex multi-subject scenes can lose detail
+- No native negative prompts (avoid via positive description)
+- Sometimes adds elements not requested ("magic")
+
+## Best Practices
+
+1. **Subject first** — describe person before setting
+2. Use natural English, story-driven prose
+3. Specify camera body + lens explicitly
+4. Use cinematography terminology, not photo hobbyist
+5. "Shot on [camera] with [lens] at f/[aperture]" works as authority anchor
+6. Include lighting setup explicitly (key, fill, rim)
+7. Reference image upload boosts consistency 3x
+8. Specify aspect ratio in prompt (16:9, 9:16, 1:1)
+9. Re-state anatomical anchors even with reference images
+10. Avoid stacked adjectives ("beautiful stunning epic")
+
+## Prompt Anatomy
+
+```
+[Subject + physical description with anchors]
++ [Action / pose / expression]
++ [Setting and environment]
++ [Lighting (specific sources, color temp)]
++ [Camera body + lens + aperture]
++ [Style notes (photorealistic, cinematic color grading, mood)]
+```
+
+## Sample Prompt
+
+```
+A 35-year-old Azerbaijani man with athletic build, short dark wavy hair,
+green eyes, charcoal wool coat. Standing by a rain-streaked window in a
+dimly lit 1960s diner, looking outside with a melancholic expression.
+Warm tungsten lighting from above creates rim on his shoulders, cool
+blue spill from the window. Shot on Sony A7 IV with 85mm Cooke S4 at
+f/1.4. Photorealistic, cinematic color grading, melancholic mood,
+shallow depth of field, 9:16 vertical aspect.
+```
+
+## Common Mistakes
+
+- **Generic adjectives** ("beautiful man") — produces generic AI output
+- **Mixing styles** ("realistic anime") — confuses model
+- **Skipping camera/lens** — flat hobbyist look
+- **Over-stacking** ("stunning, epic, beautiful, gorgeous") — degrades quality
+- **No lighting source** — flat or random lighting
+
+## Aspect Ratio Notes
+
+- 16:9 — works well
+- 9:16 — explicitly state "vertical 9:16" in prompt
+- 1:1 — works well
+- 2.39:1 — possible but specify "anamorphic 2.39:1 cinematic widescreen"
+
+## Vision Input
+
+Up to 5 reference images. Order matters — first image is primary anchor.
+
+```
+[Image 1: full body reference]
+[Image 2: front close reference]
+
+Generate this same character in NEW pose: [pose description]
+Setting: [new setting]
+Camera: [camera setup]
+IMPORTANT: maintain exact face, hair, body type from reference images.
+```
+
+## Pricing tier
+$$ — moderate cost per generation
+
+## Speed tier
+Fast — 3-5s per image

package/templates/skills/image-prompt-engineer/knowledge/models/recraft-v3.md

@@ -0,0 +1,94 @@
+# Recraft v3 — Image Model Knowledge
+
+## Provider
+Recraft
+
+## Type
+Diffusion (multi-style)
+
+## Strengths
+
+- Excellent text/typography rendering
+- Vector + raster output
+- Brand design specialty
+- Multiple style modes (realistic, illustration, vector, 3D)
+- Logo and identity work
+- Clean professional output
+
+## Weaknesses
+
+- Less photorealistic than Nano Banana for portraits
+- More design-oriented than cinematic
+- Reference image support limited
+- Style mode locks aesthetic
+
+## Best Practices
+
+1. **Choose style mode first** (realistic_image, vector_illustration, digital_illustration)
+2. Specify typography exactly when needed
+3. Use for brand/marketing visuals
+4. Be specific about color palette (hex codes work)
+5. Specify dimensions clearly
+6. Brand consistency via style mode
+
+## Prompt Anatomy
+
+```
+Style mode: [realistic_image | vector_illustration | digital_illustration]
++ [Subject and composition]
++ [Brand/color guidance]
++ [Typography if needed]
++ [Aspect ratio]
+```
+
+## Sample Prompts
+
+### Brand poster
+```
+Style: digital_illustration
+Minimalist poster for "Karavan" coffee brand. Stylized illustration
+of a single coffee cup, steam rising, against deep navy background.
+Brand color palette: navy #0A1F44, gold accent #D4A574, cream
+#F5F1E8. Typography: brand name "Karavan" in serif font at top,
+tagline "fuel for what you carry" in light sans below. 9:16
+vertical poster format.
+```
+
+### Product shot
+```
+Style: realistic_image
+Product photography of a vintage glass coffee carafe on dark wood
+table, soft window light from left, neutral background, shallow
+depth of field, premium commercial style. 1:1 square format.
+```
+
+### Vector logo
+```
+Style: vector_illustration
+Logo concept for "Karavan" coffee: stylized camel silhouette merged
+with a coffee bean shape, gold (#D4A574) on navy (#0A1F44),
+flat design, minimalist, scalable vector style. Square 1:1.
+```
+
+## Style Modes Explained
+
+- `realistic_image` — photorealistic
+- `digital_illustration` — modern illustration
+- `vector_illustration` — flat scalable graphics
+- `3d` — 3D rendered look
+- Specific sub-styles available within each
+
+## Best Use Cases
+
+- Brand identity work
+- Marketing posters
+- Social media graphics with text
+- Logo concepts
+- Product mockups
+- UI/UX inspiration
+
+## Pricing tier
+$$ — moderate
+
+## Speed tier
+Fast — 5-10s

package/templates/skills/image-prompt-engineer/knowledge/models/sd-3.5.md

@@ -0,0 +1,90 @@
+# Stable Diffusion 3.5 — Image Model Knowledge
+
+## Provider
+Stability AI
+
+## Type
+Diffusion (open weights, self-hostable)
+
+## Strengths
+
+- Open source — fully customizable
+- ControlNet ecosystem (pose, depth, edge control)
+- LoRA support (fine-tuned styles)
+- Self-hostable (no API costs after setup)
+- Negative prompts strongly supported
+- Active community models
+
+## Weaknesses
+
+- Requires technical setup (Comfy UI, A1111)
+- Quality varies dramatically by checkpoint/model
+- Less accessible to non-technical users
+- Inconsistent without proper LoRA stack
+- Default base model less polished than competitors
+
+## Best Practices
+
+1. Use weighted prompts: `(important:1.3)` or `((emphasis))`
+2. Negative prompts are essential
+3. Specify checkpoint/LoRA for desired style
+4. ControlNet for pose/composition control
+5. CFG scale 7-12 for prompt adherence
+6. Sample steps 25-50 for quality
+7. Use specific samplers (DPM++ 2M Karras for quality)
+
+## Prompt Anatomy
+
+```
+[Subject with weighted features]
++ [Action / pose]
++ [Setting]
++ [Lighting + camera]
++ [Style + quality boosters]
++ [Negative prompts]
+```
+
+## Sample Prompt
+
+```
+photo of a 35-year-old Azerbaijani man, ((athletic build)), short dark
+wavy hair, (green eyes:1.2), wearing a charcoal wool overcoat,
+standing by rain-streaked window in 1960s diner, melancholic
+expression, warm tungsten lighting from above creating (rim light:1.3),
+cool blue spill from window, shot on Sony A7 IV with Cooke S4 85mm
+at f/1.4, shallow depth of field, photorealistic, cinematic color
+grading, film grain, masterpiece, 8k quality
+
+Negative prompt: cartoon, anime, painting, illustration, drawing, 3d
+render, low quality, blurry, distorted, deformed, extra fingers,
+mutation, watermark, signature, text, amateur, snapshot, oversaturated
+```
+
+## Quality Boosters (positive)
+
+- masterpiece, best quality, 8k, sharp focus, cinematic
+- Use sparingly — too many degrades
+
+## Standard Negative Prompt
+
+For photorealism baseline:
+```
+cartoon, anime, painting, illustration, drawing, 3d render, CGI,
+low quality, worst quality, blurry, distorted, deformed, extra
+fingers, missing fingers, mutation, mutated, watermark, signature,
+text, amateur, snapshot, oversaturated, plastic skin, fake
+```
+
+## ControlNet Use Cases
+
+When advanced control needed:
+- **Pose** — match exact pose from reference
+- **Depth** — preserve composition depth
+- **Canny** — keep exact edges/structure
+- **OpenPose** — body/face landmarks
+
+## Pricing tier
+Free if self-hosted, $ if via API
+
+## Speed tier
+Variable — depends on hardware