@milenyumai/film-kit 2.0.0 → 2.0.1

package/README.md

@@ -250,6 +250,68 @@ Editorial/runtime contract shared across the package:
 - per-shot sound control uses `Audio Plan`
 - semantic consistency uses canonical `visual_world`
 
+## Reference Locking And Start/End Consistency
+
+When a reference image exists, Film-Kit now treats it as an immutable source, not a loose style hint.
+
+Default behavior across `single`, `multi`, `hybrid`, `hybrid-smart`, and `studio`:
+
+- face, hair, skin tone, body proportions, wardrobe, accessories, props, materials, patterns, and logos are treated as locked invariants
+- generated prompts should not re-describe identity details already visible in the reference
+- generated prompts should explicitly preserve the same person or object exactly as the reference
+- silent redesign, restyling, redressing, age shifting, or prop redesign is considered a failure
+
+Allowed movement inside a reference-locked shot is intentionally narrow:
+
+- pose can change only if it is explicitly declared
+- expression can change only if it is explicitly declared
+- camera or environment state can change only if it is explicitly declared
+- each start/end pair should carry only one major semantic delta unless the workflow marks a deliberate chain break or transition
+
+## Prompt Contract
+
+Start/end still prompts are now optimized for structure, not raw length.
+
+Image prompt contract:
+
+- `REFERENCE LOCK`
+- `Keep same`
+- `Change only`
+- short semantic anchor
+- targeted avoid line
+
+This means:
+
+- reference-anchored image prompts are intentionally shorter and more surgical
+- old image-quality heuristics based on minimum word counts are no longer the primary control surface
+- detailed prose is still expected for video prompts, where motion, timing, audio, and camera language need richer direction
+
+`hybrid` and `hybrid-smart` keep their compact still-prompt behavior. `single`, `multi`, and `studio` keep richer video prompt behavior for Veo, Kling, and Seedance.
+
+## Semantic QA And Render QA
+
+`visual_world` is the canonical continuity contract for semantic consistency and render review.
+
+Reference-aware runtime and reports now carry these concepts:
+
+- `identity_lock: "exact-reference"`
+- `shared_frame_invariants`
+- `allowed_change`
+- `reference_drift_forbidden: true`
+
+Generated specialist and render QA flows now explicitly fail on:
+
+- face, hair, skin, body, wardrobe, accessory, or prop drift
+- lens, angle, subject scale, light direction, or color-temperature drift
+- unexplained semantic changes between start and end frames
+- violations of declared `Keep same` invariants
+- changes outside the declared `Change only` budget
+
+Studio render QA also records render-level verdicts such as:
+
+- `reference_drift_status`
+- `start_end_contract_status`
+
 ## Seedance Runtime Coverage
 
 The package now carries Seedance-aware instructions in the generated runtime, not only in type definitions or CLI parsing.

package/build/lib/templates.js

@@ -218,11 +218,11 @@ When the user asks \`/generate\`, convert the scenario into:
 ## Output Contract
 Each \`SHOTNN.md\` is a **single file** containing ALL shot details:
 - 🔗 Chain status (FIRST / CHAINED / CHAIN BREAK)
-- İLK FRAME (start frame image prompt — min 60 words)
-- SON FRAME (end frame image prompt — min 60 words)
+- İLK FRAME (start frame image prompt — compact reference-first still contract)
+- SON FRAME (end frame image prompt — same-visual-universe + one-delta still contract)
 - AUDIO PLAN (machine-readable shot audio contract)
-- VİDEO (video prompt with audio direction — min 80 words)
-- COVERAGE SHOTS (2-3 coverage shots within same file — each min 60 words)
+- VİDEO (video prompt with audio direction — min 120 words)
+- COVERAGE SHOTS (2-3 coverage shots within same file — compact stills + detailed coverage video when needed)
 - 🇹🇷 Turkish summary for each section
 - Avoid line on EVERY prompt
 
@@ -268,7 +268,7 @@ Read .agent/VOICE-DESIGN.md when dialogue, narrator VO, or reusable speaker iden
 5. SLOW BURN: 8s default duration. Split actions into multiple shots. "Ilmek ilmek islemek."
 6. Music: NONE by default. User must explicitly request.
 7. EVERY prompt must have an Avoid line. No exceptions.
-8. Coverage shots mandatory (2-3 per main shot, min 60 words each, included in same file).
+8. Coverage shots mandatory (2-3 per main shot, included in same file; stills stay compact, coverage videos stay detailed).
 9. Frame chaining: Last frame of SHOT[N] = First frame of SHOT[N+1].
 10. Semantic consistency: \`${config.outputDir}/shot-plan.json\` must include \`visual_world\`; prompts must align camera, named movement strategy, light/shadow vector, scale, reflections, physics, anatomy risk, and contextual logic.
 11. ILK/İLK FRAME section must contain a code block even for chained shots.
@@ -334,7 +334,7 @@ All skills at: \`.agent/skills/[name]/SKILL.md\`
 - AUTO-SAFETY: Proactively reframe sensitive content
 - Frame chaining: Last frame SHOT[N] = First frame SHOT[N+1]
 - Chained ILK/İLK FRAME code block contains only \`Use SHOT[prev]_END as exact first frame\`; any new visual prompt requires CHAIN BREAK
-- Coverage: 2-3 sub-shots per main shot (min 60 words each, in same file)
+- Coverage: 2-3 sub-shots per main shot (same file; compact stills, detailed coverage videos)
 - Avoid line: MANDATORY on every prompt
 - Music: NONE by default
 - Duration: 8s default, slow burn pacing
@@ -406,7 +406,7 @@ Before generating ANY prompts, read skills from \`.agent/skills/\`:
 9. **Voice Design:** keep project-level \`voiceCast\` in \`${config.outputDir}/shot-plan.json\` and per-shot \`Audio Plan\` in each VIDEO section.
 10. **ILK/İLK FRAME:** Always include a fenced code block, even when chained.
 11. **Chained ILK/İLK FRAME:** code block contains only \`Use SHOT[prev]_END as exact first frame\`; any new visual prompt requires CHAIN BREAK.
-12. **Quality Floor:** ILK >= 80, SON >= 80, VIDEO >= 120, coverage >= 70 words.
+12. **Image Prompt Contract:** reference-anchored ILK/SON use \`REFERENCE LOCK\` + \`Keep same\` + \`Change only\`; VIDEO >= 120; coverage video >= 70.
 13. **Specificity Floor:** lens/framing, lighting, and foreground/midground/background action are mandatory.
 14. **Spatial Realism Floor:** eyeline target, plane map, shared light source, and contact/depth cues are mandatory when relational staging matters.
 15. **Semantic Consistency Floor:** \`visual_world\`, perspective/geometry, shadow vector, scale map, reflections, gravity/contact physics, anatomy risk, foreground/background coherence, contextual contradictions, and targeted semantic avoid terms are mandatory.
@@ -450,7 +450,7 @@ This workspace keeps high-level policy in \`CLAUDE.md\` and operational detail i
 - Maintain \`${config.outputDir}/shot-plan.json\` top-level \`visual_world\`
 - Keep \`Audio Plan\` blocks aligned to \`voiceCast\`
 - Keep \`ILK/İLK FRAME\` in a fenced code block even when chained
-- Quality floor and specificity floor are hard gates, not suggestions
+- Image prompt contract, specificity floor, and semantic floor are hard gates, not suggestions
 - Semantic consistency floor is a hard gate: camera/lens/camera-movement/light/shadow/scale/reflection/physics/anatomy/context must align to \`visual_world\`
 - Apply \`.agent/skills/spatial-blocking/SKILL.md\` whenever eyeline, compositing, or depth realism is critical
 
@@ -482,7 +482,7 @@ Use the Film-Kit core runtime.
 - maintain top-level \`visual_world\` inside \`${config.outputDir}/shot-plan.json\`
 - keep \`Audio Plan\` blocks valid against \`voiceCast\`
 - enforce AUTO-ANONYMOUS, AUTO-SAFETY, chaining, and coverage contracts
-- enforce quality floor: ILK >= 80, SON >= 80, VIDEO >= 120, coverage >= 70
+- enforce image prompt contract: compact reference-first ILK/SON, VIDEO >= 120, coverage video >= 70
 - enforce specificity floor: lens/framing, lighting, and foreground/midground/background action
 - enforce spatial realism: explicit eyeline target, plane map, shared light source, and contact/depth cues when needed
 - enforce semantic consistency: \`visual_world\`, perspective/geometry, shadow vector, scale map, reflection handling, physics/anatomy risk, foreground/background coherence, contextual contradictions, and scene-specific avoid terms
@@ -526,7 +526,7 @@ If using the native Claude subagent, read \`.claude/agents/prompt-engineer.md\`
    - Add machine-readable \`Audio Plan\` before every VIDEO section
    - Keep İLK FRAME as fenced code block even when chained
    - If chained, keep İLK FRAME code block to exact reuse text only; new visual prompt means CHAIN BREAK
-   - Enforce hard quality floor: ILK >= 80, SON >= 80, VIDEO >= 120, coverage >= 70
+   - Enforce image prompt contract: compact reference-first ILK/SON, VIDEO >= 120, coverage video >= 70
    - Enforce specificity floor: lens/framing + lighting + foreground/midground/background action
    - Enforce spatial realism floor: eyeline target + plane map + shared light source + contact/depth cues when applicable
    - Enforce semantic consistency floor: perspective/geometry + shadow vector + scale map + reflections + gravity/contact physics + anatomy risk + contextual contradiction check
@@ -607,7 +607,7 @@ FIRST SHOT / CHAINED from SHOT[prev]_END / CHAIN BREAK - Reason
 > Any new visual prompt in a chained ILK FRAME section requires CHAIN BREAK.
 
 \\\`\\\`\\\`
-[Image prompt — min 60 words, following prompt flow order]
+[Image prompt — compact still prompt. If refs exist: REFERENCE LOCK + Keep same + Change only + semantic anchor + Avoid]
 Avoid: blurry, low-res, noise, distorted faces, bad anatomy, extra limbs/fingers,
 plastic skin, waxy skin, on-screen text, watermark, logo, cartoon style, CGI look.
 \\\`\\\`\\\`
@@ -615,7 +615,7 @@ plastic skin, waxy skin, on-screen text, watermark, logo, cartoon style, CGI loo
 ### SON FRAME (SHOTNN_END)
 
 \\\`\\\`\\\`
-[Image prompt — min 60 words, following prompt flow order]
+[Image prompt — compact still prompt. Preserve same visual universe and declare one allowed change]
 Avoid: blurry, low-res, noise, distorted faces, bad anatomy, extra limbs/fingers,
 plastic skin, waxy skin, on-screen text, watermark, logo, cartoon style, CGI look.
 \\\`\\\`\\\`
@@ -629,7 +629,7 @@ plastic skin, waxy skin, on-screen text, watermark, logo, cartoon style, CGI loo
 ### VİDEO
 
 \\\`\\\`\\\`
-[Video prompt — min 80 words, following prompt flow order]
+[Video prompt — min 120 words, following prompt flow order]
 
 Audio direction:
 - Language: [TURKISH/ENGLISH/etc.]
@@ -654,11 +654,11 @@ inconsistent lighting, unnatural motion, on-screen text, watermark.
 🇹🇷 [Türkçe coverage özeti]
 
 \\\`\\\`\\\`
-[Coverage IMAGE prompt — min 60 words, standalone (not chained)]
+[Coverage IMAGE prompt — compact standalone still contract (not chained)]
 Avoid: blurry, low-res, noise, distorted faces, bad anatomy, on-screen text, watermark.
 \\\`\\\`\\\`
 \\\`\\\`\\\`
-[Coverage VIDEO prompt — min 60 words, with full audio direction block]
+[Coverage VIDEO prompt — min 70 words, with full audio direction block]
 
 Audio direction:
 - Language: [LANGUAGE]
@@ -812,7 +812,7 @@ Before generating ANY prompts, read skills from \`.agent/skills/\`:
 - **AUTO-SAFETY:** Proactively reframe sensitive content
 - **Frame Chaining:** Last frame of SHOT[N] = First frame of SHOT[N+1]
 - **Chain Hardening:** chained ILK/İLK FRAME code block contains only \`Use SHOT[prev]_END as exact first frame\`
-- **Coverage:** 2-3 sub-shots per main shot (in same file, min 60 words each)
+- **Coverage:** 2-3 sub-shots per main shot (in same file; stills compact, coverage videos detailed)
 - **Spatial Realism:** eyeline targets, shared light, depth scale, and anti-cutout staging must agree when subjects share frame
 - **Semantic Consistency:** \`visual_world\` controls perspective/geometry, shadow vector, scale map, reflections, physics, anatomy risk, background coherence, and contextual contradictions
 - **Avoid Line:** MANDATORY on every prompt
@@ -909,7 +909,7 @@ Before generating ANY prompts, read these skills:
 2. **Name Policy:** Dialogue naming follows \`${config.outputDir}/shot-plan.json\` policy
 3. **AUTO-SAFETY:** Proactively reframe sensitive content
 4. **Frame Chaining:** Last frame of SHOT[N] becomes first frame of SHOT[N+1]
-5. **Coverage Mandatory:** 2-3 sub-shots per main shot (in same file, min 60 words each)
+5. **Coverage Mandatory:** 2-3 sub-shots per main shot (in same file; stills compact, coverage videos detailed)
 6. **Avoid Line:** MANDATORY on every prompt (image + video + coverage)
 7. **Voice Design:** keep \`voiceCast\` in \`${config.outputDir}/shot-plan.json\` and \`Audio Plan\` in every speaking VIDEO section
 8. **Music: NONE** by default
@@ -923,10 +923,11 @@ Before generating ANY prompts, read these skills:
 
 ## Quality Floor (Hard Gate)
 Reject and regenerate any shot that fails:
-- ILK FRAME < 80 words
-- SON FRAME < 80 words
+- reference-anchored ILK FRAME missing \`REFERENCE LOCK\`, \`Keep same\`, or \`Change only\`
+- SON FRAME drifts from the same visual universe or changes more than one major beat
 - VIDEO < 120 words
-- any coverage prompt < 70 words
+- any coverage video prompt < 70 words
+- reference-visible face/hair/skin/body/wardrobe/accessories/prop design drift
 - missing explicit lens/framing/camera movement details
 - missing explicit lighting direction/intensity/atmosphere details
 - missing explicit foreground/midground/background action details
@@ -977,13 +978,15 @@ ${isKling
 - Visual prompts in English; dialogue kept in source language.
 - Main chain continuity: SHOT[N]_END -> SHOT[N+1]_START.
 - Default duration: 8s slow burn.
-- Prompt rewriter active — longer prompts (80-120 words) give more control.
+- Prompt rewriter active — longer prompts (80-120 words) give more control for VIDEO prompts; reference-anchored image prompts should stay compact and delta-only.
 
 ### Veo Image-to-Video
 When generating video from a reference image:
 - The model extracts lighting and style directly from the image
 - Do NOT repeat visual info already in the image (e.g., "warm light", "35mm film")
+- Keep the same identity, wardrobe, accessories, prop design, lens feel, light direction, and subject scale
 - Focus only on the **change**: motion, action, and audio
+- Use the same seed for same-scene reference variants and start/end pairs; change seed only for structural repair or \`CHAIN BREAK\`
 
 ### Veo Re-Take Strategy
 | Error Type | Fix | Seed |
@@ -992,7 +995,7 @@ When generating video from a reference image:
 | Extra limbs | Add "extra limbs, extra fingers" to Avoid | New seed |
 | Motion inconsistency | Simplify movement, fewer actions | Same seed |
 | Light jumps | Specify light source more precisely | Same seed |
-| Character inconsistent | Strengthen reference lock, repeat "EXACTLY" | Same seed |
+| Character inconsistent | Strengthen immutable reference lock, remove face/body/wardrobe redesign text | Same locked seed |
 | Background drift | Add background details | Same seed |
 | Robotic motion | Add "organic natural movement" + micro-behavior | New seed |
 
@@ -1038,6 +1041,9 @@ The more aligned these are, the cleaner the transition:
 - Light direction & white balance
 - Overall perspective
 
+**Keep same:** identity, face, hair, skin tone, body proportions, wardrobe, accessories, prop design, lens family, camera height, subject scale, light direction, color temperature, and location shell.
+**Change only:** one major semantic beat such as expression, hand position, body lean, prop position, or environmental state.
+
 > Large angle differences or mismatched lens feel forces the model to "guess" perspective → warp increases.
 
 **Human face/hands rules:**

package/content/agents/prompt-engineer.md

@@ -42,10 +42,10 @@ You are a senior Technical Prompt Engineer specialized in model-aware cinematic
 12. **HER prompt'ta Avoid satırı ZORUNLU (❗ coverage dahil)**
 13. **🇹🇷 Türkçe özet: Her shot ve coverage için 1 cümlelik Türkçe özet ekle**
 14. Emotional arc: Film genelinde gerilim eğrisi (1-5) uygula
-15. **Hard quality floor:** ILK >=80, SON >=80, VIDEO >=120, coverage >=70 kelime
+15. **Hard prompt floor:** reference-anchored ILK/SON use `REFERENCE LOCK` + `Keep same` + `Change only`; VIDEO >=120; coverage video >=70
 16. **Hard specificity floor:** Her promptta lens/framing, lighting ve FG/MG/BG action detayları zorunlu
 17. **Spatial realism floor:** eyeline target, plane map, shared light source, contact/weight cues ve tam ölçek derinlik mantığı gerektiğinde zorunlu
-18. **Semantic consistency floor:** `shot-plan.json.visual_world` kanonik olmalı; perspective/geometry, shadow vector, scale map, reflection handling, gravity/contact physics, anatomy risk, foreground/background coherence ve contextual contradiction her shot'ta geçmeli
+18. **Semantic consistency floor:** `shot-plan.json.visual_world` kanonik olmalı; perspective/geometry, shadow vector, scale map, reflection handling, gravity/contact physics, anatomy risk, foreground/background coherence, contextual contradiction, immutable reference lock ve allowed-change budget her shot'ta geçmeli
 
 ---
 
@@ -271,7 +271,7 @@ Before outputting ANY shot:
 - [ ] Quality floor passes? (ILK>=80, SON>=80, VIDEO>=120, coverage>=70)
 - [ ] Specificity floor passes? (lens + lighting + FG/MG/BG action)
 - [ ] Spatial realism passes? (eyeline target + plane map + shared light + contact/depth cues)
-- [ ] Semantic consistency passes? (`visual_world` fields + perspective/geometry + shadow vector + scale map + reflection/physics/anatomy/context checks)
+- [ ] Semantic consistency passes? (`visual_world` fields + perspective/geometry + shadow vector + scale map + reflection/physics/anatomy/context + reference drift + allowed-change checks)
 - [ ] Model Control block exists? (`Model`, `Preset`, `CFG`, `Transition Mode`)
 
 ### 3. Kling-Specific Gates (when model is kling-3.0)

package/content/skills/frame-chaining/SKILL.md

@@ -360,7 +360,16 @@ When preparing Start and End frames for Kling 3.0, both frames MUST share:
 | **Light direction & WB** | Same source, same color temperature |
 | **Overall perspective** | Same spatial depth and vanishing points |
 
-> Mismatched elements force Kling to "guess" → warp, melting, and artifacts increase.
+> Mismatched elements force Kling to "guess" → warp, melting, and artifacts increase.
+
+### Keep Same / Change Only Contract
+
+For a reference-locked start/end pair:
+
+- `Keep same:` identity, face, hair, skin tone, body proportions, wardrobe, accessories, prop design, lens family, camera height, subject scale, light direction, color temperature, and location shell
+- `Change only:` one major semantic beat such as expression, hand position, body lean, prop position, or environmental state
+- if two or more major semantic changes are needed, declare `CHAIN BREAK` or split into additional shots
+- do not redesign anything already visible in the reference image
 
 ### Face & Hands Stability for Start+End
 

package/content/skills/prompt-structure/SKILL.md

@@ -19,7 +19,7 @@ description: Image and video prompt templates, camera movement vocabulary, stabi
 | **Positive Framing** | Describe what SHOULD happen (negative prompts less effective) |
 | **Specific Action Verbs** | "strides" > "walks", "glides" > "moves" |
 | **Professional Terms** | Use industry-standard cinematography language |
-| **Ideal Length** | 80-120 words per video prompt |
+| **Ideal Length** | Video prompts: 80-120 words. Reference-anchored image prompts: compact 35-70 words; hybrid still pairs can go 20-50 when anchors are strong |
 | **Reference Commands First** | Always start with reference image instructions |
 | **Safety First** | Always consider filter implications |
 | **Short Sentence Rule** | Split long sentences across shots |
@@ -36,28 +36,26 @@ description: Image and video prompt templates, camera movement vocabulary, stabi
 
 | Sıra | İçerik | Örnek |
 |------|--------|-------|
-| **1** | Tek cümle sahne özeti | "Cinematic close-up of a young soldier in a dim bunker." |
-| **2** | Kim/Ne/Nerede | Karakter fiziksel tanımı, kıyafet, konum |
-| **3** | Aksiyon | Ne yapıyor, mikro-davranış, oyunculuk |
-| **4** | Kamera + Lens | "85mm f/2.0, shallow DOF, handheld" |
-| **5** | Işık + Atmosfer | "Warm oil lamp, dramatic side shadows" |
-| **6** | Audio direction | Sadece video prompt'larında |
-| **7** | Avoid line | HER prompt'ta zorunlu |
+| **1** | Reference lock / frame intent | "Same person exactly as reference. Do not modify." |
+| **2** | Keep same | Kimlik, wardrobe, prop design, lens feel, ışık yönü, subject scale |
+| **3** | Change only | Tek anlamlı değişim: ifade, el pozisyonu, aksiyon beat'i |
+| **4** | Kamera + Lens | "35mm lens feel, eye-level medium shot" |
+| **5** | Semantic anchor | Işık yönü, gölge vektörü, scale/depth, location shell |
+| **6** | Audio direction | Sadece video prompt'larında |
+| **7** | Avoid line | HER prompt'ta zorunlu |
 
 > **İlk cümle = "Bu shot ne?" sorusunun tek cümlelik cevabı olmalı.**
 
 ### Standard Template
 
 ```
-[REFERENCE LOCK section if applicable]
-
-Cinematic still frame of [subject with reference adherence] in [frozen pose].
-[Environment + time of day].
-Lighting: [specific setup].
-Camera: [framing], [lens mm], [aperture], photorealistic, crisp focus, no motion blur, no text.
-[Safety injection if needed]
-
-[Full avoid line]
+[REFERENCE LOCK] Same person/object exactly as reference. Do not modify, enhance, reinterpret, restyle, redress, or redesign anything visible in the reference.
+Keep same: identity, wardrobe/accessories, prop design, lens feel, camera height, subject scale, light direction, color temperature, location shell.
+Change only: [one intentional beat].
+[Frame intent + composition + brief action].
+Semantic consistency: [shadow direction], [scale map], [reflection/physics note].
+[Safety injection if needed]
+Avoid: [targeted failures including reference drift / outfit drift / prop drift / lens drift / light-direction drift]
 ```
 
 ### Semantic Consistency Layer
@@ -78,19 +76,27 @@ Use SHOT[prev]_END as exact first frame.
 ```
 
 Any new start-frame visual prompt must be declared as `CHAIN BREAK - [reason]`.
+
+### Reference-Anchored Image Prompt Contract
+
+When a reference exists:
+
+- do not restate face, hair, skin tone, body build, wardrobe, accessories, or prop design as freeform descriptive prose
+- use `Keep same:` for invariants and `Change only:` for one meaningful delta
+- do not re-explain style or lighting already visible in the reference image
+- use compact prompts; density matters more than length
+- if more than one major semantic change is needed, split the beat or declare `CHAIN BREAK`
 
 ### Example: Character in Environment
 
-```
-[REFERENCE LOCK] Using uploaded reference image: The soldier - Match this person's face EXACTLY. Do NOT invent new facial features.
-
-Cinematic still frame of the person from the reference, wearing dusty Ottoman artillery uniform, standing beside a massive artillery piece, one hand on the cannon barrel, determined expression. Ottoman coastal fortification, smoke-filled atmosphere, late afternoon harsh sunlight.
-Lighting: Dramatic side-light from setting sun, long shadows, dust particles visible in light beams.
-Camera: Medium shot, 35mm lens, f/4, photorealistic, crisp focus, no motion blur, no text.
-Maintaining exact likeness from the reference throughout. Historical reenactment style. Documentary approach.
-
-Avoid: blurry, low-res, noise, distorted faces, bad anatomy, extra limbs/fingers, plastic skin, waxy skin, airbrushed skin, on-screen text, watermark, logo, cartoon style, CGI look, different face than reference.
-```
+```
+[REFERENCE LOCK] Same soldier exactly as the uploaded reference. Do not modify face, hair, skin, body, uniform, fez, or cannon design.
+Keep same: identity, uniform, fez, cannon design, medium-shot 35mm lens feel, eye-level camera, late-afternoon side light, smoke-filled fortification layout.
+Change only: the soldier's expression shifts to determined while one hand rests on the cannon barrel.
+Cinematic still frame for a reaction-hold handoff.
+Semantic consistency: long shadows fall screen-right; scale map keeps soldier foreground, cannon midground, smoke depth behind.
+Avoid: distorted face, outfit drift, prop drift, lens drift, light-direction drift, bad anatomy, extra fingers, on-screen text, watermark.
+```
 
 ---
 

package/content/skills/reference-locking/SKILL.md

@@ -54,18 +54,29 @@ Replicate the artistic approach and atmosphere from the reference.
 
 ## 🔐 Reference Image = ABSOLUTE TRUTH (Text Must NOT Override)
 
-> **⚠️ KRİTİK SORUN:** Fiziksel tanım çok detaylı yazılınca Veo referans fotoğrafı ES GEÇİP kendi karakterini oluşturuyor.
-> **ÇÖZÜM:** Referans varken yüz/saç/vücut tanımı YAZMA — sadece kıyafet, yaş aralığı ve bağlam yaz.
+> **⚠️ KRİTİK SORUN:** Fiziksel tanım çok detaylı yazılınca model referans fotoğrafı ES GEÇİP kendi karakterini oluşturuyor.
+> **ÇÖZÜM:** Referans varken yüz/saç/vücut/kıyafet/aksesuar/prop tasarımı YAZMA veya yeniden yorumlama. Metin sadece korunan invariant'ı ve izinli tek değişimi söyler.
 
 > ⚠️ Reference lock sections must NEVER contain real person names. Use anonymous labels only.
 
 ### Hierarchy (Strict Priority Order)
 
-| Priority | Source | Controls | Text'te Yaz? |
-|----------|--------|----------|-------------|
-| **P0** | Reference Image | Yüz, saç, vücut, ten, göz | ❌ YAZMA — referans halleder |
-| **P1** | Prompt Text | Kıyafet, aksesuar, yaş aralığı, poz | ✅ Sadece bunları yaz |
-| **P2** | Prompt Text | Duygu, davranış, kamera, ışık | ✅ Her zaman yaz |
+| Priority | Source | Controls | Text'te Yaz? |
+|----------|--------|----------|-------------|
+| **P0** | Reference Image | Yüz, saç, vücut, ten, göz, kıyafet, aksesuar, prop tasarımı, materyal/pattern/logo | ❌ Yeniden tarif etme veya değiştirme |
+| **P1** | Prompt Text | Poz, ifade, aksiyon, kamera, çevresel beat, sahnedeki tek anlamlı değişim | ✅ Sadece izinli değişimi yaz |
+| **P2** | Prompt Text | Semantic anchor: lens, light direction, subject scale, location shell, avoid line | ✅ Her zaman yaz |
+
+### Immutable Invariants (Mandatory When Visible In Reference)
+
+If a reference clearly shows any of these, they are locked and may not drift:
+
+- face, hair, skin tone, body proportions
+- wardrobe, footwear, accessories, jewelry
+- prop design, surface texture, material, emblem, pattern, logo
+- age feel and silhouette
+
+Text may restate that these stay the same, but must not redesign them.
 
 ### ❌ Referans Varken Bunları YAZMA
 
@@ -79,17 +90,16 @@ Veo bu tanımları görünce referansı ezer ve kendi üretir:
 - Bıyık/sakal detayı ("full dark mustache")
 ```
 
-### ✅ Referans Varken Sadece Bunları Yaz
-
-```
-Bunlar referansla ÇATIŞMAZ, bağlam verir:
-- Kıyafet ("wearing dusty Ottoman artillery uniform")
-- Yaş aralığı ("early 30s") — kısa, sadece bağlam
-- Aksesuar ("red fez", "leather belt", "medals")
-- Poz/duruş ("standing tall", "leaning forward")
-- Duygu/davranış ("determined expression", "clenches jaw")
-- Aksiyon ("hoists shell onto shoulder")
-```
+### ✅ Referans Varken Sadece Bunları Yaz
+
+```
+Bunlar referansla ÇATIŞMAZ, ama sadece "aynı kalsın" veya "tek değişim bu olsun" mantığıyla yazılır:
+- Poz/duruş ("keep the same identity and wardrobe; change only the hand position")
+- Duygu/davranış ("change only the expression from neutral to determined")
+- Kamera ("keep the same lens feel and camera height")
+- Aksiyon ("the same person from the reference lifts the shell")
+- Çevresel beat ("change only the smoke density")
+```
 
 ### Doğru vs Yanlış Karşılaştırma
 
@@ -100,25 +110,24 @@ Bunlar referansla ÇATIŞMAZ, bağlam verir:
    wearing ornate Ottoman uniform with brass buttons"
   → Veo: Çok detaylı yüz tanımı → kendi veritabanından üretir → referans yok sayılır
 
-✅ DOĞRU (referans KORUNUR — metin sadece kıyafet/bağlam verir):
-  "The person from the reference, wearing ornate Ottoman uniform 
-   with brass buttons and red fez. Determined expression."
-  → Veo: Yüz tanımı yok → referans fotoğrafa bakar → aynı kişiyi kullanır
-```
+✅ DOĞRU (referans KORUNUR — metin sadece invariant + izinli değişimi söyler):
+  "[REFERENCE LOCK] Same person exactly as the reference.
+   Do not modify face, hair, skin, body, wardrobe, accessories, or prop design.
+   Keep the same uniform and fez. Change only the expression to determined."
+  → Model: Kimlik ve tasarım referansa bırakılır, sadece beat değiştirilir
+```
 
 ### Correct Combined Usage
 
 ```
-[REFERENCE LOCK] Using uploaded reference: "Commander" — 
-Match this person's face EXACTLY as shown in the reference image.
-Do NOT invent or modify any facial features.
-
-The person from the reference, wearing ornate Ottoman uniform 
-with brass buttons and red fez.
-Maintaining exact likeness from the reference throughout.
-```
-
-> **ALTIN KURAL:** Referans varken *"bu kişi nasıl görünüyor?"* sorusunu metin ile DEĞİL, referans fotoğrafıyla cevapla. Metin sadece *"ne giyiyor, ne yapıyor?"* sorularını cevaplar.
+[REFERENCE LOCK] Using uploaded reference: "Commander" — 
+Match this person EXACTLY as shown in the reference image.
+Do NOT modify, enhance, reinterpret, restyle, redress, age up/down, or improve anything visible in the reference.
+Keep the same face, hair, skin tone, body proportions, wardrobe, accessories, and prop design.
+Change only the expression to determined.
+```
+
+> **ALTIN KURAL:** Referans varken *"bu kişi/nesne nasıl görünüyor?"* sorusunu metin ile DEĞİL, referans fotoğrafıyla cevapla. Metin sadece *"aynı ne kalıyor, tek ne değişiyor?"* sorularını cevaplar.
 
 ---
 
@@ -128,14 +137,17 @@ Maintaining exact likeness from the reference throughout.
 
 > ⚠️ Use ANONYMOUS labels ("The soldier", "Commander", "The ship") — NEVER real person names.
 
-```
-[REFERENCE LOCK] Using uploaded reference images:
-- Reference 1: [Anonymous label] - Match face and body EXACTLY
-- Reference 2: [Object label] - Replicate design PRECISELY
-- Reference 3: [Style] - Match visual aesthetic
-
-[Then continue with the rest of the prompt...]
-```
+```
+[REFERENCE LOCK] Using uploaded reference images:
+- Reference 1: [Anonymous label] - Match identity EXACTLY and do not modify visible wardrobe/accessories
+- Reference 2: [Object label] - Replicate design PRECISELY; do not restyle materials/pattern/logo
+- Reference 3: [Style] - Match visual aesthetic
+
+Keep same: [identity + wardrobe + accessories + prop design + lens feel + light direction + subject scale]
+Change only: [one intentional beat]
+
+[Then continue with the rest of the prompt...]
+```
 
 ### Example with Multiple References
 
@@ -179,12 +191,13 @@ Reference reminders should appear:
 | Problem | Cause | Fix |
 |---------|-------|-----|
 | Different person appears | No explicit reference command | Add `[REFERENCE LOCK]` section at prompt start |
-| Similar but not exact | Weak reference language | Use "EXACTLY" "PRECISELY" "IDENTICAL" forcefully |
-| Reference ignored mid-prompt | Only mentioned once | Repeat reference reminder 2-3 times throughout |
-| Wrong object/prop | Generic description used | Add "from the uploaded reference, not a generic version" |
-| Costume changed | Character description too vague | Include specific clothing from reference |
-| Facial features changed | AI "improved" the face | Add "do not modify, enhance, or improve the reference face" |
-| **Reference face REPLACED** | **Text has detailed face/body description** | **REMOVE all face/hair/body text — let reference handle it** |
+| Similar but not exact | Weak reference language | Use "EXACTLY" "PRECISELY" "IDENTICAL" forcefully |
+| Reference ignored mid-prompt | Only mentioned once | Repeat reference reminder 2-3 times throughout |
+| Wrong object/prop | Generic description used | Add "from the uploaded reference, not a generic version" |
+| Costume changed | Prompt re-designed visible wardrobe | State "keep the same wardrobe/accessories exactly as reference" |
+| Facial features changed | AI "improved" the face | Add "do not modify, enhance, or improve the reference face" |
+| Accessory / prop drift | Prompt allowed reinterpretation | Lock accessory and prop design as immutable invariants |
+| **Reference face REPLACED** | **Text has detailed face/body description** | **REMOVE all face/hair/body text — let reference handle it** |
 
 ---
 
@@ -197,12 +210,12 @@ For multi-shot scenarios, maintain consistency:
 > ⚠️ Use anonymous labels — NEVER real person names in character profiles.
 
 ```
-CHARACTER: The Soldier (from reference)
-- Face: FROM REFERENCE ONLY — do not describe in text
-- Costume: Ottoman artillery soldier uniform, period-accurate
-- Accessories: Red fez, leather belt
-- Expression range: Determined, strained, victorious
-```
+CHARACTER: The Soldier (from reference)
+- Face: FROM REFERENCE ONLY — do not describe in text
+- Costume: FROM REFERENCE ONLY — keep exact wardrobe, no redesign
+- Accessories: FROM REFERENCE ONLY — keep exact accessories, no swaps
+- Expression range: Determined, strained, victorious
+```
 
 > ⚠️ Face/Build satırlarında asla metin tanımı yazma — referans fotoğrafı yeterli.
 
@@ -215,10 +228,10 @@ Each subsequent shot must include:
 
 ### 3. Track Costume Changes
 
-If costume changes (dirt, sweat, torn):
-```
-Same character from reference, now with [specific change: dust on uniform, sweat on brow]
-```
+If surface state changes without redesign (dust, sweat, wetness):
+```
+Same character from reference. Keep wardrobe design identical. Change only state: light dust on the existing uniform surface.
+```
 
 ---
 

package/content/skills/semantic-consistency/SKILL.md

@@ -28,7 +28,22 @@ Use it as the canonical source for all shots in the same scene:
     "scale_map": "foreground, midground, background object sizes and distance logic",
     "reflection_risk": "none / glass / mirror / water / metal, plus expected reflection behavior",
     "physics_constraints": "gravity, contact points, support surfaces, cloth/hair/liquid behavior",
-    "seed_strategy": "locked seed for variants / new seed for structural repair / unknown platform seed"
+    "seed_strategy": "locked seed for same-scene variants and start/end pairs / new seed for structural repair or explicit CHAIN BREAK",
+    "identity_lock": "exact-reference / n/a",
+    "shared_frame_invariants": [
+      "identity",
+      "wardrobe",
+      "accessories",
+      "prop design",
+      "lens family",
+      "camera height",
+      "subject scale",
+      "light direction",
+      "color temperature",
+      "location shell"
+    ],
+    "allowed_change": "one major semantic delta per start/end pair unless explicit CHAIN BREAK",
+    "reference_drift_forbidden": true
   }
 }
 ```
@@ -50,6 +65,9 @@ Reject or repair a prompt package if any item fails:
 - Foreground/background coherence: background does not match foreground perspective, lighting, color temperature, and style.
 - Contextual contradiction: prompt contains incompatible scene logic unless explicitly justified.
 - Scene-specific avoid line: avoid terms are generic only, missing the scene's actual failure modes.
+- Reference drift: when a reference exists, face, hair, skin tone, body proportions, wardrobe, accessories, prop design, material, pattern, or logo drift from the reference.
+- Start/end change budget: the pair changes more than one major semantic beat without an explicit `CHAIN BREAK`, or fails to declare `Keep same` and `Change only`.
+- Seed discipline: same-scene reference variants and start/end pairs fail to preserve a locked-seed strategy, or a structural repair reuses the old seed without justification.
 
 ## Chained `ILK FRAME` Rule
 
@@ -67,8 +85,11 @@ If a new `ILK FRAME` prompt is required, mark the shot as `CHAIN BREAK - [reason
 Append the relevant semantic anchors to every image prompt:
 
 ```text
+Reference lock: same person/object exactly as reference. Do not modify face, hair, skin, body proportions, wardrobe, accessories, or prop design.
+Keep same: [shared_frame_invariants].
+Change only: [allowed_change].
 Semantic consistency: [camera height], [lens family], [horizon/vanishing logic], camera movement [static or named 24-move lexicon term], single [light source] from [direction] at [angle], [color temperature], all shadows falling [direction], scale map [FG/MG/BG distances], contact points and gravity physically plausible, background perspective and lighting match the foreground, reflection handling [none/matte/accurate].
-Avoid: improper perspective, wrong scale, inconsistent shadows, impossible geometry, conflicting light sources, unrealistic reflections, floating objects, disconnected elements, broken gravity, bad anatomy, extra fingers, deformed hands, foreground-background mismatch, contextual contradiction, [scene-specific terms].
+Avoid: improper perspective, wrong scale, inconsistent shadows, impossible geometry, conflicting light sources, unrealistic reflections, floating objects, disconnected elements, broken gravity, bad anatomy, extra fingers, deformed hands, foreground-background mismatch, contextual contradiction, reference drift, outfit drift, prop drift, lens drift, unexplained semantic change, [scene-specific terms].
 ```
 
 Keep avoid lines targeted. Prefer 15-25 concrete scene-relevant terms over long universal lists.
@@ -81,6 +102,8 @@ Write `SEMANTIC-RENDER-REPORT.md` or a `Semantic Render QA` section in `RENDER-R
 
 ```markdown
 - semantic_render_status: pass/fail
+- reference_drift_status: pass/fail
+- start_end_contract_status: pass/fail
 - perspective_geometry_status: pass/fail
 - shadow_vector_status: pass/fail
 - scale_depth_status: pass/fail
@@ -91,4 +114,4 @@ Write `SEMANTIC-RENDER-REPORT.md` or a `Semantic Render QA` section in `RENDER-R
 - rerender_or_recover_actions: [none or exact SHOTNN actions]
 ```
 
-Fail the render if the actual media contradicts the canonical `visual_world` or if a chained first frame is not an exact copy of the previous rendered end frame.
+Fail the render if the actual media contradicts the canonical `visual_world`, drifts from reference-locked identity/design, violates the `Keep same` / `Change only` contract, or if a chained first frame is not an exact copy of the previous rendered end frame.