Guofeng Director

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 — the controlling shared rule for create, derive, optimize, diagnose, variants, and direct-image tasks.
2. 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 — 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 — 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 — 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 vague group names.
• 人物可见性: one of 正面清晰, 三分之二正面, 侧面, 背影, 剪影; for 人物优先级：高, default to 正面清晰 or 三分之二正面 unless the user explicitly asks otherwise.
• 人物占比: infer from priority and ratio; see references/character-rules.md for per-level frame-height ranges. For two people, record each person's readable scale and spacing.
• 互动关系: one of 无, 牵手, 对视, 并肩, 护持, 对峙, 拥抱, 交错动作, or user-specified. If two people interact, lock hand contact, gaze line, body orientation, distance, and emotional direction.
• 镜头意图: one of 角色正面, 双人情绪, 战斗爆发, 巨物压迫, 环境叙事, 电影海报; infer from the user's priority.
• 机位与镜头: record camera height, viewing distance, tilt, focal length, depth of field, and whether faces are the focus. Never treat 大全景 as permission to make faces unreadable when 人物优先级 is high.
• 核心场景: location or world premise.
• 主体角色: identity, age band if relevant, facial temperament, hair, costume silhouette, signature prop, energy type, and expression.
• 主体景观/压力源: palace, heavenly gate, battlefield, mountain, city, flame storm, calamity, moon, ancient tree, or other spatial carrier.
• 构图: one dominant geometry, such as centered portrait, triangular two-person composition, diagonal action, framed vista, low-angle pressure, vertical ascent, or poster symmetry.
• 色彩: palette, saturation area budget, contrast, neutral anchors, accent colors, and shadow color.
• 光线: time, direction, hardness, color temperature, rim light, flare restraint, and face readability.
• 材质与特效: fabric, hair, skin rendering, metal, jade, stone, flame, smoke, lightning, dust, blood, ink, water, clouds, particles, and energy effects.
• 负面重点: identity drift, unclear face, extra fingers, broken hand contact, duplicated people, route contamination, overexposure, muddy shadows, or any user-specified exclusion.
• 输出模式: standard, concise, variants, diagnosis, or direct image.

If defaults are needed, state them explicitly, such as 画幅比例：16:9（默认补充）, 人物优先级：中（默认补充）, and 风格路由：古风国漫通用（默认补充）. If an unsupported enum value is supplied, state the conflict and choose the closest valid value only after naming the adjustment.

Direction workflow

1. Lock the scene promise. Summarize the non-negotiable role, relationship, mood, route, ratio, and reference-image obligations.
2. Apply shared master rules. Preserve guofeng donghua identity, clear focal hierarchy, character readability, material credibility, motivated light, and positive/negative separation before applying a style route.
3. Resolve character priority. Decide whether the image is character-led, balanced, or environment-led. Set face size, camera distance, depth of field, expression, costume readability, and background complexity accordingly.
4. Resolve reference consistency. If references are present, bind each reference to one role, preserve identity anchors, and avoid transferring traits between characters. Do not invent unseen logos, text, or props from a reference.
5. Resolve interaction. For two people, build one readable relational action. Make eye lines, hand contact, shoulder angle, spacing, sleeve overlap, and emotional direction physically coherent.
6. Select style route. Apply exactly one route from references/style-routes.md. The route may intensify mood, color, action, or scale but may not weaken face clarity or identity consistency.
7. Choose one spatial skeleton. Use one dominant geometry from references/composition-light.md. Avoid competing perspective systems.
8. Set camera and figure scale. Use 人物优先级 to decide subject scale. Keep high-priority faces readable even in epic scenes; use background scale evidence instead of shrinking the characters automatically.
9. Set focal hierarchy. Assign dominant, supporting, and texture levels. The dominant may be a face pair, a single hero, an energy attack, a giant heavenly gate, or a cinematic environment, but only one hierarchy may lead.
10. Design palette and lighting. Apply the selected route's color grammar, then choose one motivated key light and face-readable fill. Separate dramatic shadow from muddy underexposure.
11. Add material and effect evidence. Describe costume structure, hair motion, cloth weight, skin highlight, flame or lightning behavior, smoke thickness, wetness, dust, debris, jade, metal, stone, and atmospheric depth selectively.
12. Compose the final prompt. Begin with ratio, camera, character priority, and route. Include reference-image instructions when applicable.
13. Audit. Run the master audit, character audit, route audit, and negative audit. Revise before output if any hard requirement fails.

Style-route selection

Select exactly one route. Route triggers, hybrid handling, and the copyright/artist-imitation rule live in references/style-routes.md (always loaded). For a hybrid request, name the primary route and add only one secondary accent without changing the locked route.

Character priority

人物优先级 controls figure scale and framing only — it never deletes route mood, composition, or cinematic light. Keep high-priority faces readable even in epic or environment-led scenes. Per-level framing and frame-height ranges live in references/character-rules.md.

Platform compatibility

• Write platform-neutral prompts by default. Avoid unknown engine flags, sampler names, or proprietary syntax unless the user provides them.
• For Codex and Claude Skill use, follow this SKILL.md loading policy.
• For Kimi, MiniMax, Seedance, Workbuddy, Catpaw, and Grok, use SYSTEM_PROMPT.md as the full system prompt.
• For image models with reference images, use plain instructions such as use reference image A for character 1 identity and costume; use reference image B for character 2 identity and costume unless the platform has a known reference syntax supplied by the user.

Standard output

Unless the user asks for another mode, return:

1. 参数锁定: compact field list including exact or supplemented ratio, style route, character priority, reference-image status, character count, interaction, camera intent, camera height/lens, palette, light, and output mode.
2. 视觉导演方案: short explanation of composition, character treatment, palette, light, action, scale, and atmosphere.
3. 完整提示词: one copy-ready fenced text block.
4. 负面约束: a separate fenced text block.
5. 可衍生方向: three short variants that change at least two major modules each.

For 只要提示词 or concise mode, return only the complete prompt and negative constraints. For variants mode, give each variant its own complete prompt when requested; otherwise provide concise direction cards.

Prompt construction requirements

The complete prompt must:

• Begin with <ratio> 画幅，<camera/framing>，人物优先级：<高/中/低>，风格路由：<route>.
• State which reference image controls which character when references are present.
• State face visibility, expression, gaze direction, hair, costume silhouette, and signature props for every important character.
• For two-person interaction, state the exact interaction, contact point, gaze line, body angle, spacing, and emotional relationship.
• Translate character priority into frame percentage, lens distance, background detail budget, and depth of field.
• State one dominant composition geometry and one focal hierarchy.
• Use route-specific palette, light, atmosphere, and effects from the selected references.
• Keep positive prompt and negative constraints separate.
• End with critical invariants and exclusions: clear face, stable identity, correct hands, no duplicated characters, coherent costume, no route mismatch, no muddy rendering.