@azerogluemin/ai-bootstrap 0.2.1

package/templates/skills/image-prompt-engineer/knowledge/composition.md

@@ -0,0 +1,157 @@
+# Composition — Image Generation
+
+> How subjects are arranged in frame. Critical for cinematic feel.
+
+## Core composition rules
+
+### Rule of Thirds
+Divide frame into 3x3 grid. Place subjects on intersections, not center.
+- Eyes typically on upper third line
+- Horizons on upper or lower third (not middle)
+
+### Leading Lines
+Use lines (roads, walls, light beams) to direct eye to subject.
+
+### Framing within frame
+Use doorways, windows, foreground objects to frame subject.
+
+### Negative Space
+Empty space around subject creates isolation, mood, scale.
+
+### Symmetry vs Asymmetry
+- Symmetry: formal, eerie, controlled (Wes Anderson, Kubrick)
+- Asymmetry: dynamic, natural, tension
+
+### Depth Layering
+Foreground + Midground + Background creates 3D feel.
+
+**For multi-figure / crowd scenes — MANDATORY depth distribution** (see `character-anatomy.md` "HARD RULE — Directorial Realism Density" for the full rule): never cluster all figures on one plane. Distribute across 3 layers — **FG hero figures** (2-4, full facial detail, the scene's protagonists) + **MG support figures** (2-5, contextualising roles like a vendor, a horseman, a teacher at a doorway, smaller in frame, less detail) + **BG silhouette figures** (2-6+, atmospheric "life continues" presence, near-silhouette, no face-quality concern). A real wide shot of a busy period public space typically lands at 8-14 total figures across these 3 layers; theatre/parade/bazaar scenes can exceed 50+ because BG silhouettes absorb the count with no AI face-quality penalty. Same-plane crowds (all figures at one depth) read as posed/staged and collapse face quality — always redistribute in depth.
+
+## Headroom & Lookroom
+
+- **Headroom**: space above head (not too much, not cropping)
+- **Lookroom/Noseroom**: space in direction subject is looking
+- **Leadroom**: space ahead of moving subject
+
+## Camera angles
+
+### Eye-level
+Neutral, default. Audience equal to subject.
+
+### Low angle (looking up)
+Subject powerful, dominant, heroic, threatening.
+
+### High angle (looking down)
+Subject vulnerable, small, weak, submissive.
+
+### Bird's eye / Top-down (90° from above)
+Detached, godlike, scale, geometry emphasis.
+
+### Dutch angle (tilted)
+Tension, unease, drunk, off-balance, psychological disturbance.
+
+### Worm's eye (extreme low)
+Heroic monumental, towering subject.
+
+## Framing types (recap from cinematography)
+
+- ECU, CU, MS, MWS, WS, EWS, OTS, POV, 2-shot
+
+## Aspect ratio composition tips
+
+### 16:9 (landscape standard)
+- Horizontal balance
+- Subjects positioned thirds-based
+- Headroom and lookroom standard
+
+### 9:16 (vertical for mobile)
+- Subject vertically centered or upper-third
+- Leave bottom for UI overlays (caption, music)
+- Avoid crucial detail at very top/bottom
+
+### 1:1 (square)
+- Subject centered
+- Strong negative space
+- Symmetry-friendly
+
+### 2.39:1 (cinemascope)
+- Wide horizontal real estate
+- Negative space possible
+- Anamorphic feel
+
+## Composition phrases for prompts
+
+- "Subject offset to right-third with empty leadroom on left"
+- "Symmetrical composition, subject centered"
+- "Low angle from below, subject dominates frame"
+- "Foreground bokeh frames midground subject"
+- "Dutch tilt 15° clockwise creates tension"
+- "Negative space above subject emphasizes isolation"
+
+## DON'T
+
+- Don't say "well-composed" without specifying how
+- Don't center every subject by default
+- Don't crop subjects awkwardly (mid-knee, mid-forehead)
+- Don't ignore where subject is looking
+
+---
+
+## 🔴 HARD RULE — CINEMATIC LIVED MOMENTS, never posed group photos
+
+**Every cell must feel like a LIVED MOMENT observed by a fly-on-the-wall camera from inside the era.** NEVER a posed group portrait. Even when the script reads "archive photo" or "group portrait," the underlying CELL has natural blocking, mid-action subjects, and cinematic depth — the **sepia / archive feel is applied in post** (see `color-grading.md` HARD RULE on cells in full colour).
+
+### Director's pre-prompt mindset (5 steps — mandatory before writing the cell)
+
+1. **Read the script** — what is this moment's essential meaning? What is the audience supposed to feel?
+2. **Interpret screenwriter's intent** — why this composition? What is the dramatic purpose?
+3. **Analyse the location** — how did this place actually work? What was the natural daily flow of people, light, sound? Where do bodies naturally cluster, where do they pass through?
+4. **Choose observer position** — where are WE (the camera) standing in the scene? At what eye level? Why from there and not elsewhere?
+5. **Choose mid-action moment** — what specifically is happening at this exact second? Subjects in motion verbs ("crossing," "leaning," "adjusting," "turning toward"), not static states ("standing," "posing," "looking at camera").
+
+### Natural blocking rules
+
+- **NOT** all subjects facing camera
+- **NOT** symmetric arrangement (no three-row choir formation, no row of soldiers at attention)
+- **3/4 angles preferred** over frontal
+- Some subjects **unaware of camera**; at most one or two **briefly aware** (and even then mid-glance, not posed)
+- **Imperfection is essential** — one fidgets, one adjusts clothing, one looks elsewhere, one bends down to pick something up, one is mid-yawn, one whispers to a neighbour. Natural human chaos. Real lived bodies in real lived space.
+
+### Language to AVOID in prompts
+
+- "posed for formal portrait"
+- "standing at attention"
+- "all looking at camera"
+- "symmetric arrangement"
+- "stiff stillness"
+- "lined up neatly"
+- "row of [subjects]"
+- "formal group photograph"
+
+### Language to USE in prompts
+
+- "caught mid-X" (mid-stride, mid-laugh, mid-glance, mid-pour)
+- "one looking at Y, two others at Z, the rest unaware of the camera"
+- "natural variation in posture — some leaning, some upright, some half-turned"
+- "we observe from across the courtyard / from the back of the hall / from behind a pillar"
+- "subjects unaware of the camera"
+- "asymmetric clustering — three near the window, two by the door, one alone in the foreground"
+
+### Special case — script says "archive photograph" / "old newsreel" / "B&W still"
+
+The underlying cell is still a **natural lived moment**. The sepia / monochrome / vignette / grain that sells the "archive" feel is applied in **post-production** in the video editor's LUT chain (see `color-grading.md` HARD RULE on cells in full colour). Do NOT generate the cell as a stiff posed studio portrait merely because the script labels it "archive."
+
+### Pairing with other hard rules
+
+This rule **multiplies** with two others — they only work together:
+
+- **FACE DIVERSITY** (`character-anatomy.md`) — varied blocking naturally produces varied faces; cloned faces and posed formations always co-occur
+- **PERIOD + PLACE COSTUME RESEARCH** (`character-anatomy.md`) — once you know who was actually in the room, the natural blocking writes itself (gender separation rules, social hierarchy, who clusters with whom)
+
+### Cross-references
+
+- `cinematography.md` → "Director's pre-prompt mindset" (full version with shot-list integration)
+- `character-anatomy.md` → "Face diversity" + "Period + place costume research"
+- `image-validator/knowledge/script-consistency-checks.md` — "posed vs lived" is a Qat B checkpoint
+
+**Origin:** 2026-05-21, 28-may documentary k.4 — schoolgirls cell was initially written as a stiff symmetric three-row formation posed straight at the camera. User feedback: "Burda elə bil yığılıb şəkil çəkdirirlər. Bütün kadrlarda sanki biz o dövrün içindəymişik kimi kənardan baxırıq belə olmalıdır." Rule promoted to universal toolkit knowledge.

package/templates/skills/image-prompt-engineer/knowledge/consistency-anchors.md

@@ -0,0 +1,113 @@
+# Consistency Anchors — Image Generation
+
+> The hardest problem in AI image generation: keeping the same character/location consistent across multiple shots.
+
+## The problem
+
+Each generation is independent. Prompt "young woman with red hair" produces a DIFFERENT woman every time. We need anchors to keep her recognizably the same person across 10+ generations.
+
+## Solution stack
+
+### 1. Reference images (primary mechanism)
+
+When user uploads their first generation, that image becomes the anchor. Subsequent prompts include it as **vision reference input**:
+
+```
+[Image: character_full_body.png]
+[Image: character_front_close.png]
+
+Generate the same character in a different setting:
+- Setting: hospital corridor, sterile lighting
+- Action: walking briskly toward camera
+- Camera: medium shot, 50mm at f/2.8
+- IMPORTANT: maintain exact facial features, hair, body type from references.
+```
+
+### 2. Anatomical anchor list (text-based)
+
+Even with images, prompt should re-state non-negotiable features:
+
+```
+Character: 35-year-old Azerbaijani man (refer to character sheet for
+exact face). Anatomical anchors that MUST match references:
+- Hair: short dark wavy, swept right
+- Eyes: deep green
+- Build: athletic, 5'11"
+- Distinctive: small scar above right eyebrow, slight stubble
+- Skin tone: Fitzpatrick scale 4
+```
+
+### 3. Style anchor (per project)
+
+Single style note repeated across all character generations:
+```
+Style: photorealistic cinematic, ARRI Alexa color science,
+shallow DOF, neutral exposure, 2024 contemporary realism.
+```
+
+## What can change vs what can't
+
+### MUST stay identical
+- Face shape, features, structure
+- Hair color (length/style can vary slightly)
+- Eye color
+- Body proportions
+- Skin tone
+- Scars, tattoos, glasses, distinctive marks
+
+### Can vary intentionally
+- Outfit (separate "outfit variant" prompts)
+- Pose, expression
+- Lighting and setting
+- Hair styling (tied/loose, but not color/length)
+- Time of day on subject
+
+### Should change with story
+- Age progression (over years/scenes)
+- Wardrobe based on scene
+- Emotional state via expression
+- Wet vs dry hair (weather)
+
+## Location consistency
+
+Same logic for locations:
+
+### MUST stay
+- Architecture
+- Wall colors
+- Major furniture
+- Practical light positions
+
+### Can vary
+- Time of day (sunlight angle)
+- Weather visible through windows
+- Props on tables
+- Camera angle within space
+
+## Multi-character consistency
+
+When two characters in same shot:
+1. Generate each independently first (turnaround)
+2. Use BOTH reference image sets as vision input
+3. Specify spatial relationship: "Character A on left, Character B on right, eye-line connecting"
+4. Specify both characters' features again in prompt
+
+## DON'T
+
+- Don't expect consistency without references
+- Don't change too many anchors at once
+- Don't rely only on names ("ƏLİ" doesn't anchor anything visually)
+- Don't skip the description even when sending images
+
+## Workflow recommendation
+
+1. Generate character turnaround (5 views) → pick best
+2. Upload best to library as anchor
+3. For every subsequent shot, send 1-2 best-matching references
+4. Re-state anatomical anchors in prompt text
+5. Validate: does generated character match? If not, iterate
+
+## Cross-reference
+
+- character-anatomy.md — what to describe
+- cinematography.md — how to frame for consistency

package/templates/skills/image-prompt-engineer/knowledge/general/aspect-ratios.md

@@ -0,0 +1,185 @@
+# Aspect Ratio Decision Guide — Universal
+
+> **HARD RULE — every image prompt MUST explicitly declare an aspect ratio at the end.** Never leave aspect to model default. The aspect is chosen ONCE at project intake (showrunner Addım 1B) and inherited by every downstream skill — director, storyboard-builder, image-prompt-engineer, video-prompt-engineer. Per-cell aspect drift is forbidden.
+
+This guide is **project-agnostic**. It gives the decision criteria; each project picks its aspect ratio during showrunner intake and locks it in `WORKFLOW.md` → `Aspect:` field. Every prompt then inherits that aspect — no per-cell guessing, no model defaults.
+
+---
+
+## Why this rule exists
+
+Real production incident (2026-05-21): A documentary project's first 6+ cell prompts were written at **21:9 cinemascope** by default. The actual delivery target was **YouTube + community-event projector**, which is **16:9 native**. All 6 prompts had to be rewritten or post-letterboxed. Root cause: skill defaulted to "cinematic = 21:9" without checking the delivery target.
+
+**Lesson:** Default to **16:9**, not 21:9. Cinemascope (21:9) is **cinema-only** — theatrical release, festival, prestige HDR streaming. Everything else (YouTube, TV, web, projector, social) is 16:9 native and renders 21:9 as letterboxed bars.
+
+---
+
+## Decision table — delivery target → aspect ratio
+
+| Delivery target | Aspect ratio | Why this aspect |
+|---|---|---|
+| **YouTube standard video** | **16:9 widescreen** | Native player aspect. 21:9 gets letterboxed (black bars top + bottom). |
+| **TV broadcast** (any region) | **16:9 widescreen** | Industry standard since ~2009. |
+| **Web embed / corporate / event projector** | **16:9 widescreen** | All standard projectors and web players are 16:9 native. |
+| **Documentary delivered to community event / streaming** | **16:9 widescreen** | Plays correctly on every device without letterbox. |
+| **Reel / TikTok / IG Story / YouTube Shorts / vertical mobile** | **9:16 vertical** | Mobile-first social, vertical viewport. |
+| **IG feed post (square)** | **1:1 square** | IG feed native; older convention but still common. |
+| **IG portrait feed post** | **4:5 portrait** | Modern IG feed default; takes more vertical real estate than 1:1. |
+| **Theatrical cinema release** | **21:9 cinemascope** (2.39:1) | True cinema aspect; designed for theatre screens. |
+| **Film festival submission** | **21:9 cinemascope** (or 1.85:1 flat) | Festival projection standards. |
+| **Prestige streaming HDR cinema** (e.g. Netflix premium cinematic) | **21:9 cinemascope** | Cinema-grade delivery. |
+| **Magazine / print spread / editorial card** | **varies — ASK USER** | Project-specific; depends on layout. |
+| **Character anchor / 6-view turnaround reference sheet** | **1:1 square** | Multi-panel grid is an internal asset, independent of project aspect (see `character-designer` SKILL.md HARD RULE). |
+| **Character outfit variant / single-figure reference** | **inherit project aspect** | Reference is for downstream cells, so it adopts project aspect. |
+
+---
+
+## MOST PROJECTS = 16:9
+
+**Default to 16:9 unless the project explicitly states one of:**
+- Theatrical cinema release → 21:9
+- Film festival submission → 21:9 (or 1.85:1)
+- Prestige HDR streaming cinema → 21:9
+- Mobile vertical (reel/short/story) → 9:16
+- IG square feed → 1:1
+
+If unsure → ask the user **before** writing any prompt. Never assume "cinematic look = 21:9".
+
+---
+
+## Per-aspect composition implications
+
+Choosing the aspect ratio changes how every shot is framed. Same scene, different aspect = different headroom, different subject sizing, different background reveal.
+
+### 16:9 widescreen (1.78:1) — THE DEFAULT
+
+**Frame proportions:** moderate horizontal, balanced vertical.
+
+- **Wide establishing:** subject 35-50% frame width. Sky/ceiling AND ground visible. Comfortable vertical context.
+- **Portrait / face shots:** head and shoulders fit naturally. Small headroom at top. Eye-line on upper third.
+- **OTS (over-the-shoulder):** shoulder/back of head occupies meaningful vertical space. Subject sits centered-to-third.
+- **Macro / detail (page, hand, object):** subject centered; horizontal trim minimal.
+- **Lateral pan in post:** still works — crop window slides L→R inside the 16:9 frame; pan range is shorter than 21:9 but smooth.
+- **Best for:** documentary, YouTube, TV, web, projector, corporate, multi-platform.
+
+### 21:9 cinemascope (2.39:1) — CINEMA ONLY
+
+**Frame proportions:** strongly horizontal, narrow vertical.
+
+- **Wide establishing:** subject 25-40% frame width. Sky and ground COMPRESSED — vertical context minimal. Landscape rules.
+- **Portrait / face shots:** face occupies less vertical space; needs more horizontal context around it (off-center placement common). Headroom often dropped.
+- **OTS:** shoulder takes more horizontal real estate; subject pushed further from center.
+- **Macro / detail:** strong horizontal subject (book spine, weapon, line of objects) reads well; tall narrow subjects feel cramped.
+- **Lateral pan in post:** native — designed for it. Sweeping vista moves.
+- **Best for:** theatrical, festival, prestige drama, large-format cinema.
+
+### 9:16 vertical (0.5625:1) — MOBILE SOCIAL
+
+**Frame proportions:** strongly vertical, narrow horizontal.
+
+- **Wide establishing:** subject 50-70% frame width (subject dominates because horizontal context is squeezed). Vertical space lets you stack foreground + mid + background up the frame.
+- **Portrait / face shots:** face fills most of frame; close framing is the default. Headroom often increased to leave space for text overlays at top.
+- **OTS:** rare — vertical aspect doesn't suit OTS. Prefer direct close-up or POV.
+- **Macro / detail:** vertical subjects (face, body, tall objects) read well; horizontal subjects feel cropped at edges.
+- **Lateral pan in post:** awkward — vertical frame doesn't support horizontal motion. Vertical pan / tilt works better.
+- **Best for:** Reels, TikTok, YouTube Shorts, IG Story, mobile-first social.
+
+### 1:1 square (1:1) — IG FEED + REF SHEETS
+
+**Frame proportions:** equal both axes.
+
+- **Wide establishing:** subject 40-50% frame width. Symmetrical compositions work best.
+- **Portrait / face shots:** head + shoulders fit; eye-line center or upper third.
+- **OTS:** works but feels static; both subjects roughly equal weight.
+- **Multi-panel grid (6-view reference sheet):** ideal — 2x3 or 3x2 panel layouts fit naturally in 1:1.
+- **Best for:** IG feed posts, character reference sheets, symmetrical compositions.
+
+### 4:5 portrait (0.8:1) — IG PORTRAIT FEED
+
+**Frame proportions:** slightly vertical, near-square.
+
+- **Wide establishing:** similar to 1:1 but with extra vertical breathing room.
+- **Portrait / face shots:** excellent for single-figure portraits — face capture is generous, body extends down into frame.
+- **Macro:** vertical objects read well.
+- **Best for:** IG portrait feed posts, magazine portrait editorial.
+
+---
+
+## How to declare aspect ratio in a prompt
+
+Every image prompt ends with an explicit aspect declaration. Model-specific syntax:
+
+| Model | Syntax |
+|---|---|
+| **Midjourney v7** | `--ar 16:9` (parameter) |
+| **Flux Pro / SD 3.5** | `aspect_ratio: "16:9"` (API param) + sentence in prompt: "Aspect ratio: 16:9 widescreen." |
+| **Nano Banana 2 / Nano Banana Pro** | Sentence in prompt: "Aspect ratio: 16:9 widescreen." |
+| **GPT-Image-2** | `size: "1792x1024"` (API param — closest preset to declared aspect) + sentence in prompt: "Aspect ratio: 16:9 widescreen." |
+| **Imagen 4** | `aspectRatio: "16:9"` (API param) + sentence in prompt |
+| **Ideogram v3** | `aspect_ratio: "16:9"` (API param) + sentence in prompt |
+| **Recraft v3** | `size: "1820x1024"` (closest preset) + sentence in prompt |
+
+**Universal rule:** ALWAYS declare aspect in BOTH the API parameter AND a final sentence in the prompt text. Some models ignore one or the other; declaring both is the only reliable way.
+
+**Sentence template (paste verbatim at end of prompt):**
+```
+Aspect ratio: <16:9 widescreen / 21:9 cinemascope / 9:16 vertical / 1:1 square / 4:5 portrait>.
+```
+
+---
+
+## Decision workflow (apply BEFORE writing any prompt)
+
+```
+Step 1: Read WORKFLOW.md → Aspect: field
+   ↓
+Step 1a: Is the aspect field present?
+   - YES → use it. Done.
+   - NO → STOP. Ask user: "What's the delivery target?
+     (YouTube → 16:9 / Cinema → 21:9 / Reel → 9:16 /
+      IG square → 1:1 / Other → tell me)"
+   ↓
+Step 2: Write prompt with the declared aspect.
+   - End the prompt with: "Aspect ratio: <X>."
+   - Set the API parameter to match.
+   ↓
+Step 3: Verify after writing — does the aspect declaration appear in the
+   final prompt? If missing → add it before delivering.
+```
+
+---
+
+## Common errors to avoid
+
+### Error 1 — "Cinematic = 21:9" assumption
+"Cinematic" is a **visual style** (anamorphic look, controlled lighting, film stock), NOT an aspect ratio. A 16:9 documentary can absolutely be "cinematic." Do not equate cinema look with cinema aspect.
+
+### Error 2 — Defaulting to model preset
+GPT-Image-2 defaults to 1024x1024 (1:1). Midjourney defaults to 1:1. If you don't declare aspect, you get the model's preset — often wrong for the project.
+
+### Error 3 — Per-cell aspect drift
+Every cell in a project MUST use the same aspect. If Cell 1 is 16:9 and Cell 2 is 21:9, they cannot be cut together. Aspect is a project-level decision, not a per-cell decision.
+
+### Error 4 — Forgetting the API parameter
+Declaring aspect in the prompt text alone is unreliable — some models ignore prompt aspect and use the API parameter only. Always set BOTH.
+
+### Error 5 — Choosing aspect to match the location
+"This is a wide landscape, so 21:9." NO. Aspect is chosen by delivery target, not by scene content. A wide landscape in a YouTube documentary is still 16:9 — just frame the landscape to read inside 16:9.
+
+### Error 6 — Mid-project aspect change
+Once aspect is locked in `WORKFLOW.md`, all already-generated assets (character refs, location refs, cells) are tied to it. Changing aspect mid-project means re-generating everything or post-cropping (with loss). Lock the aspect at intake, not later.
+
+---
+
+## Cross-references
+
+- **Showrunner intake:** `showrunner/SKILL.md` → Addım 1B captures aspect ratio as the second project-level question.
+- **Image prompt writing:** `image-prompt-engineer/SKILL.md` → "Davranış qaydaları" → Aspect ratio HARD RULE.
+- **Storyboard cell inheritance:** `storyboard-builder/SKILL.md` → every cell prompt inherits project aspect, no per-cell override.
+- **Director shot composition:** `director/SKILL.md` → framing decisions adapt to declared aspect.
+- **Image validation:** `image-validator/knowledge/prompt-consistency-checks.md` → F-16 checks aspect ratio match against project standard.
+- **Character reference sheet exception:** `character-designer/SKILL.md` → 6-view turnaround stays 1:1 regardless of project aspect.
+
+---
+
+*Version: 1.0 | Created: 2026-05-21 | Trigger: real production incident — project defaulted to 21:9 when delivery was 16:9 YouTube/event.*

package/templates/skills/image-prompt-engineer/knowledge/general/prompt-budget.md

@@ -0,0 +1,167 @@
+# Prompt Length Budget — Universal Compression Strategy
+
+> **HARD RULE — image generation promptları char budget-ə uyğun olmalıdır.** Verbose 8,000-9,000 char promptlar AI image modellərində ya **silent quality drop** ya da **explicit "prompt too long" error** verir. Bu fayl model-aqnostik kompresiya strategiyasıdır — `models/gpt-image-2.md` və `models/nano-banana.md` HARD RULE bölmələri ilə birlikdə oxunmalıdır.
+
+---
+
+## Niyə char budget məcburidir
+
+AI image modelləri **early-token bias** ilə işləyir:
+
+- Promptun **ilk ~500-800 tokeni** dominant olur
+- ~500 tokendən sonra instruction-ların effektivliyi **kəskin düşür**
+- 9,000 char promptda kritik detallar (refs, identity locks, anti-grit, lighting, anti-synthetic) **sondadırsa**, model onları **görməyəcək**
+- Texniki "hard limit" ≠ effektiv limit
+- Web UI-lərdə isə hətta soft limit də 2,000-4,000 char-dadır
+
+---
+
+## Verified char limits (web research, 2026-05-21)
+
+| Model | Hard limit | Praktik effektiv | Web UI cap |
+|---|---|---|---|
+| GPT-Image-2 (OpenAI) | 32,000 char | **~3,000 char (~500 söz)** | dəyişir |
+| Nano Banana 2 (Gemini 3.1 Flash Image) | 131K tokens API | ~8,000 char API | **2,000 char web** |
+| Nano Banana Pro (Gemini 3 Pro Image) | 65K tokens API | ~8,000 char API | 2,000 char web |
+| Midjourney v7 | 6,000 char | ~3,500 char | 6,000 char |
+| Flux Pro | ~512 tokens optimal | ~2,000 char | dəyişir |
+
+**Universal target (parallel multi-model istifadədə):**
+- **Hədəf:** ≤2,800 char per prompt
+- **Maksimum:** 3,000 char (mütləq aşma)
+- **Səbəb:** Hər iki modelə eyni prompt yapışdırılır → ən kiçik effektiv limit (~3,000) hökm edir
+
+---
+
+## Detection thresholds (`wc -c` verification habit)
+
+Hər prompt yazandan sonra **gözlə ölçmə, hesabla**:
+
+```bash
+cat prompt.txt | wc -c
+```
+
+| Char count | Verdikt | Action |
+|---|---|---|
+| **≤2,500** | ✅ optimal | Davam et |
+| **2,500-3,000** | ⚠️ pass | Aggressive compression mümkündür |
+| **3,000-4,000** | ⚠️ risk | GPT-Image-2-də quality drop, kompress et |
+| **>4,000** | ❌ critical | Aggressive cut LAZIMDIR |
+| **>8,000** | ❌ blocker | Multiple model error riski, mütləq yenidən yaz |
+
+---
+
+## How to apply (compression strategy)
+
+### 1. CUT aggressive — bu sektorları SİL və ya KOMPRESS et
+
+- ❌ **Verbose color hex palette (~500 char)** — modeldə **rəng adı** kifayətdir, hex lazım deyil
+- ❌ **Lengthy "NOTHING ELSE" lists** — 1 cümləyə sıxış
+- ❌ **Verbose lighting breakdown** ("KEY: ... FILL: ... BACKGROUND: ...") — 2-3 cümlə bəs
+- ❌ **Material palette section** — material adlarını subject description-da inline yaz
+- ❌ **Extensive lamp/prop anatomy paragraph** — 1-2 cümlə (clear glass chimney + flame + brass base)
+- ❌ **Multi-line camera/grade block** — 1 cümlə (ARRI ALEXA 35 + Cooke S7/i 50mm + Kodak Vision3 5219)
+- ❌ **Aspect ratio repetition** — 1 dəfə kifayətdir
+- ❌ **Repeated identity recap** — bir cümlədə anchor instruction + bir cümlədə action (təkrar yox)
+
+### 2. KEEP these (kritik info, kompromiss yox)
+
+- ✅ **Subject identity** + age + ethnicity + critical wardrobe (1-2 cümlə)
+- ✅ **ATTACH/anchor reference instruction** ("preserve face from image 1")
+- ✅ **Camera angle + lens + DOF** (1 cümlə)
+- ✅ **Primary action / mid-action moment** (1 cümlə)
+- ✅ **Lived-moment / fly-on-the-wall clause** (1 cümlə)
+- ✅ **Lighting source + temperature** (1 cümlə)
+- ✅ **Period-specific HARD RULE flags relevant to THIS kadr** (lamp anatomy IF lamp visible, anti-grit IF hands visible, face diversity IF multi-figure, flag IF flag visible)
+- ✅ **Full color + post-grade clause** (1 cümlə)
+- ✅ **Anti-synthetic clause** (1 short sentence: "Avoid photorealistic, HDR, AI art aesthetic — film frame look")
+- ✅ **Aspect ratio** (4-5 words)
+
+### 3. Template (≤2,800 char skeleton)
+
+```
+[SHOT TYPE + CAMERA ANGLE + SUBJECT IDENTITY with anchor preserve instruction in 2-3 sentences]
+
+[PRIMARY ACTION + LIVED MOMENT framing in 1-2 sentences]
+
+[WARDROBE + period-specific HARD RULE flags relevant to this kadr in 2-3 sentences]
+
+[SETTING + BACKGROUND in 2-3 sentences]
+
+[LIGHTING source + temperature + key direction in 1-2 sentences. If lamp visible: 1-sentence period-pure anatomy.]
+
+[Lens + DOF + camera grade: 1 sentence with ARRI ALEXA 35 + Cooke S7/i + Kodak Vision3 5219]
+
+[Full color, sepia/grade in post: 1 sentence]
+
+[Anti-synthetic + anti-AI-art: 1 short sentence]
+
+[Aspect ratio: 21:9 cinemascope]
+```
+
+---
+
+## Concrete example — compression demonstration
+
+### ❌ V1 (9,168 char — TOO LONG, GPT-Image-2 "prompt too long" error)
+
+```
+PRIMARY SUBJECT — An over-the-shoulder (OTS) cinematic shot, looking PAST the right shoulder of a seated/standing 38-year-old man (the SAME PERSON FROM IMAGE 1, attached as anchor reference: preserve EXACTLY his face structure, dark full beard, dark hair, complexion — even though only the back/side of his head, shoulder, and hands are visible in this shot, identity must match the anchor for video editing continuity). We see him from behind and slightly to his right, the camera positioned just over his right shoulder, looking down at what he is holding.
+
+[...8,000 more chars of verbose layout / materials / colour palette hex codes / per-light breakdown / lamp anatomy / repeated identity recap / multi-line camera spec...]
+```
+
+**Failure mode:** Model ya error atır, ya ilk 500 token-dən sonra hər şeyi ignor edir (identity lock, anti-grit, lighting timeline sondadırsa itir).
+
+### ✅ V2 (2,713 char — IDEAL, eyni kritik info, effektiv limit içində)
+
+```
+Over-the-shoulder cinematic shot looking past the right shoulder of a 38-year-old man (same face as IMAGE 1, preserve identity exactly). Top of frame: back-right of his dark fur hat. Upper-right: his dark wool jacket shoulder (soft defocus). Lower foreground: his clean lean hands holding a newspaper (no ink stains — historical intellectual, anti-period-grit rule).
+
+The newspaper fills 60% of frame, held at ~15° angle, masthead-up. [Period-specific masthead + body text description in 2-3 sentences.]
+
+Background soft bokeh defocus: dark cast-iron platen press at right, wooden type cases at left, kerosene lamp glow lower-left edge, pressman silhouette far back. Aged warm-cream plaster walls.
+
+Setting: [period + place] evening — PRE-electric era. The lamp at lower-left edge is a period-pure KEROSENE lamp: brass fuel reservoir base, brass wick-knob, clear glass bulbous chimney with visible yellow-orange flame inside (~2700K). NO frosted shade, NO electric bulb.
+
+Key light: warm kerosene lamp from camera-left foreground casting raking light across newspaper page. Subtle fill from second off-frame oil lamp. Background ambient very dim.
+
+Lived moment, fly-on-the-wall observer behind him — NOT a posed prop shot.
+
+Camera: 50mm lens equivalent on ARRI ALEXA 35 + Cooke S7/i at T/2.0 (newspaper masthead and upper body text sharp, foreground hands/shoulder slightly soft, background dreamy bokeh). Kodak Vision3 5219 emulation, warm amber-ochre period grade, ~15% desaturated, lifted filmic blacks, visible 500T grain, mild vignette.
+
+Full color period-drama register (sepia applied in DaVinci post, NEVER baked here). Avoid photorealistic, HDR, magazine aesthetic, AI art look — this is a Barry-Lyndon-style film frame.
+
+Aspect ratio: 21:9 cinemascope.
+```
+
+**Səbəb V2 işləyir:**
+- Eyni kritik info (identity lock, lighting, period anatomy, anti-grit, anti-synthetic, aspect)
+- GPT-Image-2 limit-i içində (~3,000 char altında)
+- Nano Banana web UI yaxınlığında (2,000-2,800)
+- Bütün important instruction-lar **ilk 500 token-də** və ya yaxın yerdə
+- Heç bir repeated info, heç bir verbose hex palette, heç bir multi-line lighting breakdown
+
+---
+
+## Workflow integration
+
+Bu rule **bütün AI creator layihələrində** tətbiq olunur — universal AI image generation limit reality. Knowledge tree-də yer alır:
+
+- `models/gpt-image-2.md` — GPT-Image-2 üçün model-spesifik HARD RULE
+- `models/nano-banana.md` — Nano Banana API vs web UI fərqi
+- `general/prompt-budget.md` — bu fayl, universal compression strategy
+- `SKILL.md` "Davranış qaydaları" — `wc -c` verification habit reminder
+- `image-validator/knowledge/prompt-consistency-checks.md` — F-15 checkpoint (validator side)
+
+---
+
+## Related rules
+
+- `models/midjourney-v7.md` — MJ 6,000 char hard limit, ~3,500 char effective
+- `models/flux-pro.md` — Flux ~512 token sweet spot
+- Prompt complexity budget (max ~1 primary zone + 1 secondary + 5-7 materials + 2 lights) — ayrı qayda, bu char budget-dən fərqlidir, lakin paralel tətbiq olunur
+
+---
+
+*Versiya: 1.0 | Yaradılma: 2026-05-21 (real production incident-dən sonra: 9,168 char prompt GPT-Image-2-də "prompt too long" error verdi, web research ilə model-spesifik limitlər təsdiq olundu).*