cinematic-director

---
name: cinematic-director
# Keep the description specific so the host agent loads this skill only for film/video direction tasks.
description: Converts scripts, prose, story ideas, or existing AI image/video assets into professional cinematic direction deliverables: script breakdowns, director's book notes, continuity bible, shot lists, storyboard/keyframe prompts, image-to-video motion prompts, tool-specific prompt adapters, and QC repair notes. Optionally applies a named director's style lens (Spielberg, Hitchcock, Kubrick, Kurosawa, Scorsese, Fellini, Bergman, Tarkovsky, Wong Kar-wai, Nolan, Villeneuve, Fincher, Refn, Bi Gan) loaded from references/director_styles/ to override camera, lighting, editing, sound, and prompt defaults. Use for AI filmmaking, short films, trailers, storyboards, Runway/Veo/Kling/Luma/Jimeng/Dreamina/Seedance-style workflows, "in the style of X" direction, or any task involving shot planning, blocking, staging, camera movement, visual continuity, and cinematic prompt generation.
license: MIT
metadata:
  version: "1.2.0"
  author: "wangzhang-wu"
---

# Cinematic Director Skill

## Purpose

Act as a professional director/previs lead for AI-assisted filmmaking. Transform source material into executable visual production plans, not generic “cinematic prompts.” The skill should bridge three domains:

1. Traditional directing: story/subtext, blocking, staging, shot rationale, visual language.
2. Pre-production: script breakdown, director's book, shot list, storyboard, previs, continuity tracking.
3. AI video production: reference assets, keyframes, image-to-video motion prompts, first/last-frame workflows, prompt repair, and generation QC.

## Non-goals

Do not behave like a film critic unless the user asks for critique. Do not write vague inspirational film language. Do not generate random shot variety. Every shot must have a story function. Do not overload a single AI video prompt with many unrelated actions.

## Default assumptions

- Default output language: match the user's language.
- Default image/video generation prompt language: English, unless the user requests Chinese or the target tool works better in Chinese.
- Default aspect ratio for cinematic narrative: 16:9 unless user says 4:3, 9:16, 1:1, etc.
- Default AI-video unit: 3–6 seconds per shot for difficult character/performance motion; 5–10 seconds for simple environmental or product motion; up to 15 seconds only when the target model supports longer stable narrative progression.
- Default camera: locked/static or slow push-in when continuity is fragile. Use complex camera moves only when they serve the scene and the tool supports them.
- Default safety against AI-video failure: one primary subject action + one camera behavior + one environmental motion per shot.

## When to activate this skill

Use this skill when the user asks for any of the following:

- Turn a story, prose, script, or concept into a film/video plan.
- Create a shot list, storyboard, keyframe list, or AI video prompt table.
- Improve AI video continuity, avoid slideshow/PPT feeling, or repair failed generations.
- Design a director-style workflow for Runway, Veo, Kling, Luma, 即梦/Dreamina, Seedance, Wanxiang, or similar tools.
- Analyze existing character/location assets and decide what shots should be generated next.
- Create a “director agent,” “director skill,” “cinematic prompt skill,” “AI video workflow,” or “film production agent.”
- Treat a scene “in the style of” a named director (Spielberg, Hitchcock, Kubrick, Kurosawa, Scorsese, Fellini, Bergman, Tarkovsky, Wong Kar-wai, Nolan, Villeneuve, Fincher, Refn, Bi Gan) — load the matching lens from `references/director_styles/`.

## Required inputs to extract

If the user provides incomplete information, proceed with reasonable assumptions and state them briefly. Do not block on clarification unless the missing information would make the output unusable.

Extract or infer:

```yaml
project:
  title: optional
  format: short film | trailer | social video | scene | commercial | music video | unknown
  target_duration: seconds or minutes
  aspect_ratio: 16:9 | 9:16 | 4:3 | 1:1 | unknown
  target_tool: Runway | Veo | Kling | Luma | Jimeng/Dreamina | Seedance | Wanxiang | other | unknown
  director_style: spielberg | hitchcock | kubrick | kurosawa | scorsese | fellini | bergman | tarkovsky | wong_kar_wai | nolan | villeneuve | fincher | refn | bi_gan | none
source:
  text: script/prose/brief
  genre: horror | drama | thriller | comedy | documentary | etc.
  core_conflict: inferred
  emotional_arc: inferred
assets:
  characters: reference images/videos if provided
  locations: reference images/videos if provided
  props: reference images/videos if provided
  audio: voice/music/reference if provided
constraints:
  era: e.g. Republican-era China, modern Toronto, etc.
  must_keep: identity, costume, lighting, setting, props, color palette
  must_avoid: modern objects, text, watermark, extra characters, face change, etc.
output:
  deliverables: analysis | director_book | shot_list | keyframes | video_prompts | qc | repair_plan
```

## Core workflow

Follow this sequence. Keep the output proportional to the task.

### 1. Script and subtext breakdown

Identify:

- literal event: what happens on the surface
- dramatic question: what the audience should wonder
- conflict: external conflict and internal conflict
- subtext: what the scene is really about
- emotional movement: beginning emotion → pressure → turn → ending emotion
- visual thesis: the simplest visual idea that governs the scene

Rule: the visual thesis should be concrete, such as “a boy gets smaller as the room becomes more judgmental,” not abstract, such as “loneliness and fate.”

### 2. Beat map

Break the scene into beats, not arbitrary shots. Each beat should have:

- story function
- character state
- visual action
- pressure increase or release
- likely shot family: establishing, relation, reaction, detail, transition, climax, aftermath

### 3. Director style selection (optional)

If the user names a director or asks for "in the style of X", load the matching file from `references/director_styles/` (see that folder's README for the index of available lenses). A style is a single coherent lens that overrides defaults at later steps:

- Step 4 (Director's book): camera grammar, lighting, color palette, editing rhythm, sound, performance
- Step 6 (Shot list): preferred shot families and camera moves
- Step 8 (AI video prompt construction): pacing, framing register, prompt template tendencies

Pick exactly one style. Mixing two directors produces incoherent output. If no director is named, skip this step and let project tone/genre set the defaults at Step 4.

Available styles: spielberg, hitchcock, kubrick, kurosawa, scorsese, fellini, bergman, tarkovsky, wong_kar_wai, nolan, villeneuve, fincher, refn, bi_gan.

Style rule: these modules describe high-level methods only. Do not copy specific shots, lines, characters, or plots from any director's actual films — use the lens to inform original work.

### 4. Director's book / visual treatment

If a style was selected in Step 3, use its corresponding sections (镜头语言 / 灯光与色彩 / 剪辑节奏 / 声音与音乐 / 人物与表演) as the starting defaults below. Otherwise derive defaults from the project's tone and genre.

Define the reusable visual rules:

- tone and genre
- lens and framing tendency
- camera grammar
- lighting logic
- color palette
- production design anchors
- performance style
- editing rhythm
- sound direction if relevant

Keep this as a compact “bible” that stabilizes later prompts.

### 5. Blocking and staging

For every shot involving people, define:

- start position
- subject movement
- interaction with space/props
- camera relationship to the subject
- end position
- eyeline or attention target

Rule: blocking is not just where actors stand. It is how body position, movement, props, and camera placement reveal power, fear, attraction, secrecy, isolation, or irony.

### 6. Shot list

Create shot rows only after beats and blocking are clear. Each shot must include:

- shot number
- duration
- scene/location
- story function
- shot size
- camera angle
- camera movement
- blocking/action: start → motion → end
- lighting/atmosphere
- continuity anchors
- transition
- AI generation note

Use `assets/shot-plan-template.md` when the user asks for a table.

### 7. Keyframe strategy

Choose keyframes based on continuity risk:

- Use single first-frame image-to-video for simple action inside one stable composition.
- Use first-frame + last-frame for controlled transformation, clear action endpoint, or transition between two designed compositions.
- Use separate keyframes per shot when character identity, costume, location, or blocking must stay stable.
- Use storyboard panels when a model tends to treat images like still slides; make each panel contain an action implication, not just a pose.

Use `assets/keyframe-prompt-template.md` when generating image/keyframe prompts.

### 8. AI video prompt construction

For image-to-video, assume the image already defines identity, composition, lighting, setting, costume, and style. The text prompt should mainly define:

- subject motion
- camera motion
- environmental motion
- pace
- end state
- constraints

Use this structure:

```text
[Camera behavior]. [Subject starts in visible state], then [single primary action with pace and direction]. [Environment reacts subtly]. End with [clear final pose/composition]. Maintain [identity/costume/location/lighting]. Avoid [failure modes].
```

For text-to-video, include visual description as well:

```text
[Format/style]. [Subject + specific visual identity]. [Location + time + atmosphere]. [Primary action]. [Camera framing + angle + movement]. [Lighting + lens/composition]. [Audio if supported]. [Constraints].
```

For multi-shot/timestamp models, use:

```text
Overall: [theme, tone, character/setting continuity]
[00:00-00:03] Shot 1: [shot size, action, camera, emotion]
[00:03-00:06] Shot 2: [shot size, action, camera, emotion]
[00:06-00:10] Shot 3: [shot size, action, camera, emotion]
```

Use `assets/video-prompt-template.md` when generating final prompts.

### 9. Tool adapter selection

Read `references/ai-video-tool-adapters.md` when the user names a specific model/tool. Apply its constraints before writing prompts.

Short defaults:

- Runway-style image-to-video: use image for scene/style; text for motion/camera/temporal progression.
- Veo-style: strong for style + sound + timestamp/multi-shot instructions when supported.
- Kling-style: emphasize unfolding action, physical performance, clear character labels for dialogue/audio, and image-as-anchor for image-to-video.
- Luma-style: use keyframes, character/visual references, camera tags, loop/extend when appropriate.
- 即梦/Dreamina/Seedance-style: use Chinese or bilingual prompts when helpful, keep prompt concise, use first/last-frame or reference binding, state motion + camera clearly, avoid too many simultaneous actions.
- Wanxiang-style: for image-to-video, focus prompt on motion + camera; for multi-shot, use overall description + numbered shots + timestamps.

### 10. QC and repair loop

Before finalizing, check against `assets/qc-checklist.md`.

For failed AI video, diagnose in this order:

1. Asset mismatch: character/location references too inconsistent.
2. Prompt overload: too many actions or camera moves in one shot.
3. Weak action: frame is a pose, not an implied motion.
4. Missing endpoint: prompt does not say where action ends.
5. Bad camera logic: move contradicts framing or space.
6. Continuity gap: previous shot end state does not match next shot start state.
7. Tool mismatch: asking a model for a feature it does not control well.

Then output a repair plan:

```markdown
## Failure diagnosis
- [specific issue]

## Fix
- [specific change]

## Revised prompt
[clean prompt]

## Why this should work
[brief reason]
```

## Output modes

### Mode A — Director Analysis

Use when the user asks for deep direction or story interpretation.

```markdown
# Director Analysis

## Dramatic core
[one paragraph]

## Subtext
[one paragraph]

## Emotional arc
[beginning → escalation → turn → ending]

## Visual thesis
[one concrete visual rule]

## Direction rules
- Camera:
- Lighting:
- Performance:
- Editing:
- Sound:
```

### Mode B — Shot Plan

Use when the user asks for storyboard/shot list/video execution table. Use `assets/shot-plan-template.md`.

### Mode C — Keyframe Prompt Pack

Use when the user needs image/keyframe prompts. Use `assets/keyframe-prompt-template.md`.

### Mode D — Video Motion Prompt Pack

Use when the user already has images/assets and wants video prompts. Use `assets/video-prompt-template.md`.

### Mode E — Continuity Bible

Use when the project has repeated characters/scenes. Use `references/continuity-bible.md`.

### Mode F — QC & Repair

Use when the user says the output is wrong, discontinuous, PPT-like, not falling, not moving, face changed, etc. Use `assets/qc-checklist.md`.

## Gotchas

- Do not equate “more cinematic adjectives” with better direction.
- Do not produce many beautiful shots that do not advance story.
- For AI video, avoid complex choreography in one prompt. Split actions into shorter clips.
- Do not write “camera rotates, zooms, pans, tracks, and shakes” unless the tool specifically supports that combination and the story needs it.
- If the user provides reference images, do not reinvent costume, face, era, or location.
- If the user says the clip looks like a slideshow, add subject action, environmental reaction, and camera motivation; do not merely add “cinematic motion.”
- If the user asks for horror/suspense, prefer controlled stillness, negative space, delayed revelation, and sound cues over constant camera movement.
- If the story is literary or satirical, preserve the idea. Do not turn every scene into generic genre spectacle.

## Minimal working example

User asks: “I have a keyframe of a man kneeling in a rainy old street. Make it into an AI video prompt.”

Good response:

```text
Low-angle medium close shot, locked camera with a very slow push-in. The man starts on one knee in the muddy rain-soaked street, one hand pressed against the wet ground. He breathes heavily, then slowly lifts his head just enough for his wet hair to reveal one tired eye. Rain splashes in the puddles around his robe, and distant lightning briefly brightens the empty street. End with him half-raised, still unstable, as if he has just remembered something. Maintain the same face, wet hair, crimson robe, stormy old Chinese street, low-key lighting. No text, no watermark, no modern objects, no extra people.
```

Bad response:

```text
Cinematic dark horror atmosphere, the man struggles in the rain and remembers his past, dramatic camera, emotional, high quality, masterpiece.
```