guofeng-director

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: guofeng-director
description: Generate, derive, optimize, and diagnose structured AI image prompts for Chinese guofeng donghua visuals, including dark xianxia / Renegade Immortal-like cold oppression, Battle Through the Heavens-like hot-blooded flame action, and general elegant guofeng animation. Use for character-forward ancient Chinese fantasy prompts, clear frontal faces, character consistency from reference images, two-person interactions such as hand-holding or eye contact, cinematic composition, lighting, palette, negative constraints, prompt rewrites, variants, or prompt audits across Codex, Claude, Grok, Kimi, MiniMax, Seedance, Workbuddy, Catpaw, and similar image-prompt workflows.
---

# Guofeng Director

Turn a short character or scene idea into a coherent, auditable guofeng donghua visual direction. Preserve character clarity and emotional intent while routing the scene through one controlled visual style.

## Loading policy

Load only the references a task needs. Do not read every file for every request.

**Always load (all modes):**

1. [references/master-rules.md](references/master-rules.md) — the controlling shared rule for create, derive, optimize, diagnose, variants, and direct-image tasks.
2. [references/style-routes.md](references/style-routes.md) — every task applies exactly one style route: `逆仙黑暗`, `斗破热血`, or `古风国漫通用`.

**Load on demand (only when the trigger applies):**

3. [references/character-rules.md](references/character-rules.md) — whenever any person, face, identity, reference image, costume, expression, pose, or two-person interaction is present or implied. Skip for environment-only scenes with no identity-critical figures.
4. [references/composition-light.md](references/composition-light.md) — whenever designing or changing ratio, camera, figure size, composition, palette, lighting, action, atmosphere, or cinematic rendering. A full structured prompt normally needs this; a pure wording-level optimize may not.
5. [references/negative.md](references/negative.md) — before writing final negative constraints or diagnosing failures. Any output that includes a `负面约束` block needs this.

For **diagnose** tasks, load the reference matching the reported symptom (e.g. identity drift → character-rules; muddy shadows or route contamination → composition-light and negative) rather than all four subsystem files.

## Operating modes

Infer one mode from the request:

- **Create**: Expand a short concept into a complete structured prompt.
- **Derive**: Retain character identity, style route, or visual DNA while changing setting, pose, mood, palette, lens, ratio, or action.
- **Optimize**: Rewrite an existing prompt for clearer hierarchy, role consistency, controllable camera, and model-friendly wording.
- **Diagnose**: Identify likely causes of unclear faces, identity drift, broken hands, weak interaction, dull lighting, clutter, weak scale, generic fantasy styling, or route mismatch, then provide a corrected prompt.
- **Variants**: Produce meaningfully different directions. Change at least two major modules per variant; do not merely swap adjectives.
- **Direct image**: Invoke an available image-generation capability only when the user explicitly asks to generate an image or directly output the picture. A bare parameter block means prompt generation, not image generation.

## Parameter lock

Record explicit inputs before directing the scene. Do not silently replace them.

- `画幅比例`: one of `16:9`, `21:9`, `4:3`, `3:2`, `4:5`, `9:16`; supplement `16:9` when omitted.
- `风格路由`: one of `逆仙黑暗`, `斗破热血`, `古风国漫通用`; infer from subject and emotional keywords when omitted, otherwise default to `古风国漫通用`.
- `人物优先级`: one of `高`, `中`, `低`; default to `中` when people are central, `低` when the request is environment-led.
- `参考图状态`: one of `无参考图`, `单人参考图`, `双人参考图`, `多参考图`; record which identity, costume, hairstyle, color, or prop each reference controls.
- `人物数量`: `0`, `1`, `2`, or `多人`; record role labels instead of vag