🎬 TipAi Cinema Skills

AI Video Creation Skill Suite

Derived from TipAi desktop TPEMA engine · Agent-loadable · Zero dependencies

中文文档

---

🧠 Architecture (v5 — Self-Contained)

  User pastes script + says "Start Creating"
         │
         ▼
  skills/ai-video-studio.md  ← 🎯 Single self-contained file (load this one)
         │
         ├── Step 1-2: Script Analysis + Character Lock-in (all templates inline)
         ├── Step 3:   Shot Breakdown (shot sizes / camera moves / transitions inline)
         ├── Step 4:   Per-Shot Generation (10-dimension parameter grid inline)
         ├── Step 5:   Expression Injection (25+ AU + physiology models inline)
         └── Step 6:   Cross-Audit (5-layer audit + 20 quality checks inline)
               │
               ▼
         ✅ Audit passed → deliver / ❌ Auto-fix → re-audit
               │
               ▼
         validator/check-shots.js  (code-layer additional validation)

v5 Core Principle: Load one file, get everything needed at runtime. No external references.
Skill = Runtime instructions for AI, with mandatory checkpoints, no skipping allowed
Validator = Real executable code that verifies output consistency
📚 Reference Manuals = Sub-skill files preserved for deep reference, not needed at AI runtime

---

🔬 Extreme Detail Mode (v5 default)

Standard vs Extreme — same shot SHOT_001 prompt comparison:

Standard:  "ECU of hands around cup. Warm amber light, soft bokeh. Static." (15 words)

Extreme:   "ECU. LIGHTING: 4200K key@45°/30° china_silk(grid_60), key:fill 3:1,
           rim sep 45°, shadow 0.3. CAMERA: ARRI_Alexa_35 S35, UltraPrime_85mm @ T1.4,
           ISO 800, shutter 180°, tripod, H130cm D40cm. CHAR: HAIR #1a1a1a 25cm
           straight flyaway0.2; SKIN #f5f0e8 fine_pores SSS[cheek,nose,ear];
           EYES iris#4a2810 ct_window_rect lash_medium; HANDS nail2mm #c48a7c.
           FACS: AU1=0.05 AU2=0.02 AU4=0.05... (23 values). ACTION: thumbTrace
           0.3cm/s arc, steamWobble 0.8Hz. ATMOSPHERE: steam+dust dens0.15
           rising-swirling, vol0.6. COLOR: Kodak2383 LUT, grain2.5%, sat
           R1.05 G0.95 B0.90. POST: fine_grain 2.5%, halation0.1, CA0.3px.
           ✅ Cross-audit passed → DELIVER" (140+ words, all 10 dimensions filled)

Information density: 50x improvement. All descriptions upgraded from label-level to physics-level.

10 Dimensions

Dimension	Parameter Examples
D1 Lighting Physics	4200K, 45°/30°, china_silk, 3:1, shadow 0.3
D2 Camera Physics	ARRI Alexa 35, 85mm @ T1.4, ISO 800, 180°
D3 Character Anatomy	Hair/skin/eyes/hands/clothing hex values + physics params
D4 FACS Expression	23 AU × intensity + blink/pupil/breathing/head pose
D5 Micro-Movement	Speed cm/s, trajectory, frequency Hz, amplitude cm, weight feel
D6 Atmosphere	Particle type+density, convection, volumetric intensity
D7 Color Science	LUT, grain%, saturation curves, contrast, warm/cool balance
D8 Material Properties	Roughness/specularity/SSS/fresnel/metallic
D9 Audio-Visual Sync	Sound source position, reverb, viseme sequence, ambient
D10 Post-Production	Grain type+%, halation, CA, vignette, gate weave

5-Layer Cross-Audit

Generate → Audit1(Self-Consistency) → Audit2(Dimension Coverage) → Audit3(Physics Plausibility)
         → Audit4(Reference Integrity) → Audit5(Synthesizability) → ✅Deliver / ❌Fix→Re-audit

---

🎯 Core Skill (v5 Self-Contained)

📂 ai-video-studio.md — Load this one file, that's it.

Script Analysis → Character Lock-in → Shot Breakdown → Per-Shot 10-Dim Generation → FACS Injection → 5-Layer Cross-Audit

v5 Features

Feature	Description
🔒 User-Controlled Entry/Exit	AI never auto-activates. User says "Start Creating" to enter, "End Creation" to exit
📋 6 Mandatory Checkpoints	Each step outputs [CHECKPOINT_N], user confirms before next step
🚫 Anti-Laziness Rules	No abbreviations, no skipping, no placeholders, no partial completion
📦 Fully Self-Contained	10-dim params, AU tables, audit rules — all inline in one file
🔄 Cross-Turn State Tracking	Progress state block output each turn, AI never loses track

Built-in Capabilities (all inline, no other files needed)

Module	Description
🔍 Script Analysis	Extract characters (appearance/personality/traits), scenes (lighting/tone/time), emotional arc
🔒 Character Lock-in	Invariant traits (≥3) across shots + micro-anatomy profile (hair/skin/eyes/hands/clothing)
📋 Context Table	Cross-shot state tracking: position → emotion → lighting → timeline → transition continuity
✂️ Shot Breakdown	5 shot sizes + 7 camera moves + 7 transition types + 5 breakdown rules
🔬 10-Dimension Generation	D1-D10 full physics-level parameter injection
😊 Expression Injection	Punctuation → 25+ AU mapping + emotion formulas + physiology (blink/pupil/breath/head/gaze)
✅ 5-Layer Cross-Audit	Self-consistency → Dimension coverage → Physics plausibility → Reference integrity → Synthesizability

---

🛡️ Validator

node validator/check-shots.js examples/context-table.json --project examples/project.json

# Options:
#   --json      JSON format output
#   --verbose   Show rule descriptions + passed shots

  Quality Report
  ══════════════
  Total shots: 4
  ✅ Passed: 4
  ⚠ Warnings: 0
  ❌ Errors: 0
  Pass rate: 100%

20 Rules

#	Rule	Severity
1	Shot ID uniqueness	Error
2	Character reference validity	Error
3	Scene reference validity	Error
4	Duration reasonability	Warning/Error
5	Shot type repetition ≤3	Warning
6	Prompt element completeness	Warning
7	Adjacent shot transition defined	Warning
8	Character state continuity	Warning
9	Same-scene light delta ≤±3°	Warning
10	Scene switch transition	Warning
11	Character key features + anatomy cross-shot consistency	Error
12	Same-scene spatial continuity	Warning
13	Emotional continuity	Warning
14	Rhythm pacing (tension scenes)	Warning
15	Version format (semver)	Warning
16	Referential integrity (dangling refs)	Error
17	Micro-detail completeness (non-label-level)	Warning
18	FACS AU vector injection for dialogue shots	Warning
19	Cross-audit marker presence	Info
20	10-dimension full coverage check	Warning

---

📖 Usage

Scenario: User pastes a story, wants to create an AI short film

Step 1: Load Skill

Copy the contents of skills/ai-video-studio.md to your AI agent, or tell it:

"Load the skill at https://github.com/aitippro/tipai-cinema-skills/blob/main/skills/ai-video-studio.md"

Step 2: Say the trigger word and paste script

Start Creating. Analyze this script:

She pushed open the wooden cafe door, the brass bell chiming softly.
Afternoon sunlight streamed in diagonally, gilding every table gold.
She scanned the room — the corner seat was still empty,
the one where they'd first met.

She took a deep breath and walked toward the corner.
Her fingertips brushed the tabletops along the way, as if touching memories.
As she sat down, her eyes glistened,
but a faint smile tugged at the corner of her lips.

Outside, a figure stopped in his tracks.
His fingers hovered above the brass bell, unable to push.

Step 3: AI executes the 6-step workflow

The AI follows the Skill instructions step by step:

Step	What the AI does	Output
1. Script Analysis	Extract characters, scenes, emotional arc	Character table + Scene table
2. Character Lock-in	Build character anatomy profiles	[CHAR_01] locked description
3. Shot Breakdown	Segment into shot sequence	Numbered shot list
4. Prompt Generation	Output AI video prompt per shot	English prompt (10-dim format)
5. Expression Injection	Inject punctuation → AU mapping for dialogue shots	FACS expression data
6. Cross-Audit	Cross-check characters/scenes/lighting	Audit report

Step 4: Confirm each checkpoint

At each [CHECKPOINT_N], the AI pauses and waits. You say:

• "Continue" — advance to next step
• "Modify [item]" — go back and fix
• "Pause" — save state
• "End Creation" — exit

Step 5: Validate output (optional)

Save the AI-generated JSON as my-shots.json and run:

node validator/check-shots.js my-shots.json --project my-project.json

---

🚀 Quick Start

git clone git@github.com:aitippro/tipai-cinema-skills.git

1. Load skills/ai-video-studio.md into your AI agent
2. Say "Start Creating" and paste your script
3. The AI executes the 6-step workflow with [CHECKPOINT_N] at each step
4. Say "Continue" to advance, "Pause" to save state, "End Creation" to exit
5. Optional: Run node validator/check-shots.js to validate output quality

---

📂 Directory

tipai-cinema-skills/
├── skills/
│   ├── ai-video-studio.md          ← 🎯 Main Skill v5 — self-contained, load this one
│   ├── micro-detail-injection.md   ← 📚 Reference: 10-dimension parameter engine
│   ├── cross-audit.md              ← 📚 Reference: 5-layer cross-audit
│   ├── director-storyboard.md      ← 📚 Reference: storyboard engine
│   ├── multimodal-prompt.md        ← 📚 Reference: multimodal prompt (text-to-image + image-to-text)
│   ├── prompt-lifecycle.md         ← 📚 Reference: lifecycle & unified schema
│   └── tpema-expression.md         ← 📚 Reference: TPEMA expression engine
├── validator/
│   └── check-shots.js              ← 🛡️ 20 quality rules (code-layer)
├── .github/workflows/
│   └── test.yml                    ← 🤖 CI auto-validation
├── examples/
│   ├── project.json                ← Sample project (2 characters, 2 scenes)
│   ├── shots-example.json          ← Standard mode sample shots
│   ├── context-table.json          ← Context table sample (with transitions/states)
│   └── extreme-shot.json           ← Extreme detail single shot sample
├── README.md                       ← 📖 English documentation (this file)
├── README_CN.md                    ← 📖 中文文档
└── LICENSE

---

📜 Origin

Extracted and refined from the TipAi v1.0 desktop application core engine:

TipAi is a local-first, full-stack AI prompt engineering desktop tool React 19 · Electron 41 · Rust NAPI-RS · AES-256-GCM github.com/aitippro/TipAi

---

_{MIT License · TipAi Team © 2026}