SKILL.md · diff
git:20260425.9d4c44d to git:20260425.318222b
45 added, 22 removed. Audit C to C.
# SKILL.md
```yaml
name: reborn-master-prompt-generator
version: 2.2.0
description: |
Meta-skill that generates a niche-locked master prompt for AI video creation
- given a NICHE and a reference YOUTUBE_URL. Runs Phase 0 (niche-type classifier)
- + 5-agent reverse-engineering pipeline (Niche Intelligence → Video Analysis →
- Pattern Engine → Prompt Architect → Output Compiler) and emits a single
- paste-and-run master prompt conforming to a 34-section canonical structure
- including a SCRIPT WRITING SYSTEM section with mode-based narration
- (full / minimal / text-only / silent), register lock, hook formula, retention
- mechanics, and cue notation. v2.2 adds 4 strict quality gates (variant safety,
- hook-resolution traceability, locale-native register completeness, audio-visual
- coherence). The output prompt is topic-agnostic within the niche and reusable
- across an entire channel.
+ given a NICHE and a reference YOUTUBE_URL. Runs a 5-agent reverse-engineering
+ pipeline (Niche Intelligence → Video Analysis → Pattern Engine → Prompt
+ Architect → Output Compiler) and emits a single paste-and-run master prompt
+ conforming to a 35-section canonical structure including SCRIPT WRITING
+ SYSTEM (mode-based narration), CHARACTER / SUBJECT CONTINUITY (tool-specific
+ lock methods), valid SCENE JSON schema with worked example, and enforced
+ numbering consistency. The output prompt is topic-agnostic within the niche
+ and reusable across an entire channel.
inputs:
- NICHE (required)
- YOUTUBE_URL (required)
- TARGET_TOOL (optional, default VEO3)
- DURATION (optional, default longform)
- LANGUAGE (optional, default English)
- EXTRA_NOTES (optional)
outputs:
- - master_prompt (markdown, 1800+ words, 34 sections)
+ - master_prompt (markdown, 1500+ words, 33 sections)
entrypoint: META-PROMPT.md
license: MIT
```
---
## How to invoke
1. Open `META-PROMPT.md`. Copy the entire file.
2. Paste into Claude / GPT-5 / Gemini.
3. Append your INPUT block at the bottom.
4. Submit. Receive the master prompt.
- ## Canonical output sections (34)
+ ## Canonical output sections (35)
The generated master prompt MUST contain, in order:
1. Header line (`You are an AI system specialized in...`)
2. PRIMARY OBJECTIVE (PRIORITY ORDER)
3. INPUT SYSTEM
4. DESIGN AUTONOMY RULE
5. HIGH RETENTION SYSTEM
6. VISUAL HOOK RULE
7. VARIATION ENGINE
8. SECTION GENERATION SYSTEM
9. STRUCTURE / WORLD CONSISTENCY
10. SPATIAL / TEMPORAL CONTINUITY
11. CAMERA SYSTEM
12. CONTINUITY RULE
13. NO REGRESSION RULE
14. OBJECT / ELEMENT PERSISTENCE
15. {NICHE_PROCESS} STAGES
16. MICRO-STAGE DETAIL SYSTEM
17. {KEY_COMPLETION} RULE
18. SYMMETRY / BALANCE RULE
19. STAGE RULES
20. PROGRESS RULE
21. {NICHE_KEY_MOMENT} RULE
22. TRANSITION SYSTEM
23. PACING
24. AUDIO SYSTEM
- 25. **SCRIPT WRITING SYSTEM** (always present; MODE = full-narration / minimal-narration / text-only / silent)
+ 25. SCRIPT WRITING SYSTEM (always present; MODE = full-narration / minimal-narration / text-only / silent)
26. LIGHTING SYSTEM
27. COLOR GRADING LOCK
28. IMAGE QUALITY SYSTEM
- 29. MASTER IMAGE TEMPLATE
- 30. SCENE SYSTEM (JSON schema)
- 31. OUTPUT CONTROL
- 32. FAILSAFE
- 33. STYLE LOCK
- 34. END OF PROMPT marker
+ 29. MASTER IMAGE TEMPLATE (now includes `Character Lock:` field)
+ 30. **CHARACTER / SUBJECT CONTINUITY** (always present; method = midjourney_cref / flux_lora / kling_first_last_frame / runway_references / sora_storyboard / veo_image_condition / none)
+ 31. SCENE SYSTEM (JSON schema with valid placeholder syntax + worked example)
+ 32. OUTPUT CONTROL
+ 33. FAILSAFE
+ 34. STYLE LOCK
+ 35. END OF PROMPT marker
+ ## Numbering consistency rule (CRITICAL)
+
+ Every emitted master prompt picks ONE numbering scheme and applies it everywhere:
+ - **Scheme A (Section-aligned, 1-indexed):** Section 1 = cold open, Section N = coda. Used for fixed-arc niches (history, lore, mystery).
+ - **Scheme B (Process-aligned, 0-indexed):** Stage 0 = before state, Stage N = final reveal. Used for transformation niches (restoration, craft, exploration).
+ - HIGH RETENTION SYSTEM, STAGES, PACING, AUDIO, LIGHTING, and SCRIPT all reference the same scheme. No mixed indexing.
+
+ ## Character continuity (CRITICAL when named subjects exist)
+
+ If the niche features named figures or recurring focal objects, the master prompt MUST emit tool-specific lock syntax:
+ - Midjourney v7: `--cref <URL> --cw 100`
+ - Flux 1.1 Pro: `<lora:name_v1:0.85>`
+ - Kling 2.5: first-frame pinning
+ - Runway Gen-4: `@reference_name` tags
+ - Sora 2: storyboard character cards
+ - VEO 3.1: `image_condition` parameter
+
+ If no named subjects exist, the section still appears with `method: none`.
+
## Quality gates
- Variation engine: ≥ 4 axes × ≥ 5 variants
- Stage progression: ≥ 6 niche-specific stages
+ - Numbering consistent across all sections (no mixed Section/Stage indexing)
- Audio system: maps SFX + music + VO to every stage
- - **Script Writing System: explicit MODE, full register lock, second-by-second hook formula, ≥ 4 retention mechanics, cue notation reference, closing formula, ≥ 6 anti-fail rules, worked output example**
+ - Script Writing System: explicit MODE, full register lock, second-by-second hook formula, ≥ 4 retention mechanics, cue notation reference, closing formula, ≥ 6 anti-fail rules, worked output example
+ - Character / Subject Continuity: explicit method + tool-specific syntax in MASTER IMAGE TEMPLATE
+ - Scene JSON: valid syntax (parses as JSON), filled fields, worked example with real values
- Lighting system: progresses across stages
- - FAILSAFE: ≥ 8 niche-specific rules
+ - FAILSAFE: ≥ 8 niche-specific rules including character lock failures (if applicable)
- Style Lock: focal length + grade + tones specified
- - Length: ≥ 1800 words (script section adds ~300 words to baseline)
+ - Length: ≥ 1800 words
- Zero unfilled `{placeholders}` in final output
## Special commands (post-generation)
- `show phase 1..5` — emit the internal agent output for that phase
- `variant` — regenerate only the VARIATION ENGINE section with different axes/values
- `tighten` — reduce output by 30–40% (keep rules, drop examples)
- `loosen` — expand examples and edge cases by 30–40%
- `script only` — emit only the SCRIPT WRITING SYSTEM section
- `flip mode to {full-narration|minimal-narration|text-only|silent}` — regenerate SCRIPT WRITING SYSTEM with that mode
+ - `flip numbering` — switch between Scheme A (Section-aligned) and Scheme B (Process-aligned)
+ - `flip character lock to {midjourney_cref|flux_lora|kling_first_last_frame|runway_references|sora_storyboard|veo_image_condition|none}` — regenerate CHARACTER CONTINUITY section with that method
## Safety / scope
- Does not generate scripts, images, or videos itself — only the master prompt.
- Does not fetch external URLs autonomously beyond what the host model supports. If URL fetch fails, will request manual paste of title + transcript + first-30s frame description.
- No fabricated observations: if the reference video cannot be analyzed, the agent halts and asks the user.