Communitygithub.com

jerry00007/wuxia-director

Complete AI martial arts cinema direction system. Use when creating wuxia/xianxia fight scene prompts, storyboards, or films with AI video generation. Covers the full pipeline: character design, fight choreography, camera language, VFX layering, and quality assurance. Validated with real production data.

What is wuxia-director?

wuxia-director is a Claude Code agent skill that complete AI martial arts cinema direction system. Use when creating wuxia/xianxia fight scene prompts, storyboards, or films with AI video generation. Covers the full pipeline: character design, fight choreography, camera language, VFX layering, and quality assurance. Validated with real production data.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/jerry00007/ai-wuxia-director/tree/HEAD

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

AI Wuxia Director · Complete Methodology

Consolidated from 5 days of production, 6 completed films, 5 frame-by-frame analyses, 4 open-source platform reverse-engineering sessions, and 2 A/B regression tests.

Pipeline: 7 Steps From Story to Film {#pipeline}

① SPEC      → Duration (15s primary), aspect ratio (9:16), intensity, speed tier
② FRAMEWORK → Match scene type to battle skeleton (see references/02)
③ LEDGER    → Write combat ledger: each round must change at least one outcome
④ DRAFT     → Write prompt in Reverse-Engineering Format (see references/04)
⑤ LIGHTING  → Select and paste one lighting preset module (see templates/lighting-presets.md)
⑥ GATE      → Pass all 4 quality gates (see below)
⑦ DELIVER   → Generate → download → edit at 1.35x → publish

The 12 Iron Rules {#rules}

Every rule validated by at least one A/B test or frame-by-frame analysis. Violating any rule = rework.

Combat & Action

Rule 1 — Segment or Fail One generation = one action process (attack / counter / collision). Never ask AI to render multi-strike choreography in a single generation. The model will compress excess actions until invisible.

❌ "Three rapid strikes exchanged"
✅ Segment A: strike + block
   Segment B: counter + dodge
   Segment C: final clash + impact wave
   → Edit segments at 1.35x speed

Rule 2 — Two Events Maximum Per Shot Each timed block contains at most 2 events. Slow motion appears once per film (the decisive moment).

Rule 3 — State Chain + Secondary Reaction Every shot: start state → single action (trigger/contact/result) → end state plus one secondary reaction (cloak fluttering, dust settling, embers fading).

Visual Control

Rule 4 — First-Frame Occupancy Lock The first visible frame must contain all required characters in correct positions. No empty establishing shots. No delayed character reveals. No abstract fillers.

Rule 5 — Measurable Spatial Language Write positions as measurements, not vague words.

❌ "near the pillar" / "dozens of soldiers"
✅ "within 1 meter of the pillar" / "soldiers fill both flanks and depth, 3 in sharp foreground focus"

Rule 6 — Physics Lock Enforce: weight transfer, cloth/hair delay, weapon mass (wrist angle reacts to mass, acceleration/deceleration). Prohibit: floating bodies, weightless weapons, frictionless feet, teleporting.

Camera & Rhythm

Rule 7 — Six-Item Camera Specification Every shot specifies: scene space / starting frame / character blocking / camera movement / lens & focus / endpoint & rhythm. One camera movement per shot.

Rule 8 — Emotion as Lens Behavior Translate emotions into observable physical actions:

❌ "he looks angry"
✅ "pupils contract with reflected blade light / fist tightens, veins rise on hand / smile freezes then shatters, eyes redden"

Aesthetic Standards

Rule 9 — Eastern Beauty Anchor Protagonists must be young (20s), East Asian featured (flat brow ridge, shallow eye sockets, monolid/narrow phoenix eyes, black irises, soft jawline). Explicitly state "Chinese/Eastern facial features" or the model drifts to Western deep-set faces.

Rule 10 — Material Collision Visual interest comes from contrasting materials (silk × metal, rough cloth × polished blade), not from color variety. One dominant color + one accent color per film.

Generation Discipline

Rule 11 — Sell the Hit Prioritize environmental impact waves (snow spray/petal burst/cracked ground). For body contact: hide the contact point (camera behind attacker's shoulder) + show the reaction (head snaps in force direction) + sound + aftermath.

Rule 12 — Sound Timeline Sound starts on the first frame. Pre-strike: ambient silence. Impact: peak loudness. Post-strike: one-beat silence. Closing: half-beat fade.

The 4 Quality Gates {#gates}

Every prompt must pass ALL gates before generation. Every film must pass before publication.

Gate 1 — Technical QA (Objective)

  • Fingers normal? Weapons unbent?
  • No face morphing across shots?
  • No body clipping / teleporting?
  • Fallen characters remain visible?
  • No AI-generated text/watermarks?

Gate 2 — Silent Self-QA (CINEDANCE)

  • All @tags actually used?
  • Props in correct hands?
  • Actions physically possible?
  • Time blocks consistent?
  • First frame correct?
  • Every character's position/gaze clear?
  • Lighting not flat?
  • QA hidden from output?

Gate 3 — Aesthetic Judgment (Subjective)

  • Eastern face anchor present?
  • Material collision (silk × metal)?
  • One dominant color + one accent?
  • Hair-light rim glow?
  • Bokeh background (architecture/candles)?
  • User has final veto

Gate 4 — Platform Compliance

  • Violence softened (weapons shatter, not people)
  • No blood/gore keywords
  • No copyrighted character names
  • AI-generated content label checked
  • Platform-safe language only

Reverse-Engineering Format (Master Template)

【人物固有资产】(Character Fixed Assets — repeat verbatim every segment)
Character N: <age, ethnicity, face details (Eastern anchor), eyes, hair, 
clothing materials, weapons, quality descriptors>

【固定锚点】(Fixed Anchors — repeat verbatim every segment)
<Scene, sky, ground, color system, rendering quality>

【光线布置】(Lighting Setup — paste from presets, one only)
<Light direction, quality, falloff, ambient, mood>

0-1.5s: <extreme close-up: eye/pupil reflecting threat, sound>
1.5-4s: <pull back to wide shot: spatial reveal, threat density, sound builds>
4-9s: <combat escalation: segmented actions, energy trails, impact>
9-12s: <decisive moment: THE one slow-motion, color inversion, maximum VFX>
12-15s: <hero freeze: silhouette, aftermath, particles, half-beat fade>

【交接状态】<Final frame state = next segment's starting state>
【声音时间线】<Second-by-second sound events>

Negatives: no text/watermarks, normal proportions, no extra movements

Model Strategy

TierCost (15s)QualityUse For
Premium (Seedance 2.5)300 credits (¥13.5)9/10 actionFinal production
Standard (Seedance 2.0 mini)90 credits (¥4)6/10Drafts, testing

Budget circuit breaker: Max 2 re-rolls per shot. On 3rd failure, revise prompt instead. Monthly ceiling: 2000 credits.

Asset Management

  • Every character gets a three-view sheet (front/side/back) before any video generation
  • Assets named: @char_<name>_v<version> — referenced by ID, never by filename
  • Registry file tracks all assets with usage logs and deprecation status
  • Style consistency: same style suffix + same character reference image across ALL shots in a film

Fight Frameworks (Quick Reference)

TypeSkeletonBest For
One vs ManyWide establish → staggered attacks → use-enemy-as-shield → keep fallen visible → count survivors1vN crowd fights
DuelTrading blows → momentum shifts → weapon lock → decisive strike1v1 climactic battles
Ranged EnergyPower-up → projectile exchange → shield/deflect → ultimate beamMagic/cultivation battles
Solo BreakthroughRoute establish → obstacle 1 → obstacle 2 → final barrier → escape/stopChases, escapes

The VFX Density Layer

Visual effects are not garnish — they are the main course. Every action beat needs:

MomentVFX Requirement
Power-upEnergy cracks on weapon, air distortion, floating debris
MovementEnergy trail (comet tail), speed lines, environmental disruption
ImpactCross/ring shockwave, particle burst, ground fissure with glow
Bullet-timeFrozen particles, reflected light on suspended debris, hair suspended
FinishChain reaction (mountain collapse), lingering energy, slow-dissipating particles

References

  • references/01-camera-language.md — 12 movements with prompts and verification checks
  • references/02-fight-choreography.md — Combat ledger, frameworks, density tables
  • references/03-lighting-color.md — 8 presets, color theory, director style cards
  • references/04-prompt-format.md — Full reverse-engineering format specification
  • references/05-vfx-layer.md — VFX density rules and templates
  • references/06-production-pipeline.md — CLI automation, budget, asset registry

Templates

  • templates/character-sheet.md — Character design card prompt
  • templates/storyboard-format.md — Shot-by-shot template
  • templates/scene-library.md — 10 wuxia scenes with complete prompts
  • templates/emotion-vocabulary.md — 80 emotions as physical behaviors
  • templates/lighting-presets.md — 8 ready-to-paste modules
  • templates/fight-frameworks.md — 4 battle type skeletons

Examples

  • examples/01-sword-cold-nineteen-states.md — Full production: "One sword chills 19 provinces"
  • examples/02-moonlit-twins.md — Full production: identity reveal duel
  • examples/film-analysis-template.md — Automated QA analysis report template

Related Skills