Director

Orchestrate AI video production for any style — talking head, b-roll, product demo, podcast, or custom. This skill chains the pipeline steps. Each step has its own skill with detailed instructions.

Inputs

Ask the user for:

1. What kind of video? — style, tone, purpose (e.g. "lofi b-roll surfing video", "talking head explaining a product", "podcast clip")
2. Character — directory with character.json containing name, promptBase, referenceSheet.url, optionally voiceId
3. Scenes — scene prompts from /prompt-writer, or describe what happens and this skill will write basic prompts. For best results, use /prompt-writer first — it applies model-specific prompt structure and camera vocabulary.
4. Model — Seedance 2, Sora 2, or Kling 3.0 (see /model-provider)

Determine Format Config

Based on what the user described, determine these settings:

Setting	When ON	When OFF
Sound	Video has spoken dialogue	Silent / overlay-only / music-only
Speech gate	Video has spoken dialogue	No dialogue
STS voice swap	Video has spoken dialogue and character has voiceId	No dialogue or no voice clone
Loudnorm	Video has spoken dialogue	Silent
Overlay text	User provides overlay text	No overlay needed
Music	User provides music prompt or URL	No music
Motion in prompts	B-roll, lifestyle, action content	Talking head (stillness required)

Pipeline

Run these steps in order. Each step is a separate skill — invoke it or follow its instructions inline.

0. /prompt-writer (recommended)

Write model-ready scene prompts from the video concept. Applies model-specific structure, camera vocabulary, and constraints. Skip if the user already has scene prompts.

1. /director-frame-gen

Generate start frames (and optionally end frames) for each scene using the character reference sheet.

2. /director-clip-gen

Generate video clips from the frames. Configure sound on/off based on format config.

3. /director-qa

Run QA gates on each clip — anatomy check, critic check, and optionally speech verification. Retry failed scenes (max 5 attempts). If a scene fails all retries, abort.

4. /editor-post-production

Assemble the final video — concat, loudnorm, STS, overlay, music, edit, resize, upload. Only runs the steps that apply based on format config.

Key Behaviors

• Adapt to any style — the pipeline is the same, only the config changes
• Ask before assuming format — if unclear whether sound/overlay/STS is needed, ask
• Abort early — if a scene fails 5 retries at any gate, stop. Don't burn credits.
• Report per-scene results — critic scores, attempt counts, final CDN URL