video-director
Use whenever the user asks you to make, plan or produce a whole video - a launch teaser, promo, explainer, trailer, social clip or short film - from an idea, a brief, a changelog or data. Start here before animating anything; brief_gate.py fixes length, aspect and voice-over budget, then it runs beat board, code animatic, optional AI shots and conform to a finished MP4.
Try your first task
A source-grounded starter task has not been verified for this skill yet.
SKILL.md
name: video-directordescription: Use whenever the user asks you to make, plan or produce a whole video - a launch teaser, promo, explainer, trailer, social clip or short film - from an idea, a brief, a changelog or data. Start here before animating anything; brief_gate.py fixes length, aspect and voice-over budget, then it runs beat board, code animatic, optional AI shots and conform to a finished MP4.argument-hint: "[brief.json | one-sentence goal]"context: forklicense: MITmetadata: version: 1.0.0 build_pattern: "Previz-to-final pipeline: deterministic brief gate, beat timing, seekable HTML animatic rendered frame-exact with Playwright, ffmpeg conform with loudness QC. Generation MCPs are optional and only touch shots the user approved and paid for." distinct_from: "demo-video (screenshots to product walkthrough); video-content-strategist (channel strategy, hooks, SEO); cinematic-shot-prompting (camera prompts for one generated clip)"Video Director
Portability: every step up to the final MP4 runs locally. Prerequisites: Python 3, ffmpeg on PATH, and Playwright (Python or Node) with its Chromium build. No account, no credits. Generated footage is an optional, opt-in upgrade per shot.
You direct short films the way previz studios do: lock the story, timing and framing in a cheap animatic first, and spend money only on shots that survive review. The animatic is code (a seekable HTML timeline), so it renders the same every time and a one-word change re-renders one shot, not the film.
Workflow
P=${CLAUDE_PLUGIN_ROOT}/skills
# 1 Brief: derive aspect + VO word budget; ask what is open; refuse what is unsafe
python3 $P/video-director/scripts/brief_gate.py --brief film/brief.json --out film/brief.json
# 2 Beats → timed scene (refuses beats that do not fit the length)
python3 $P/video-beat-board/scripts/beat_board.py --brief film/brief.json --beats film/beats.json --out film/scene.json
# 3 Lint, compile, preview, render
python3 $P/video-animatic/scripts/motion_linter.py --scene film/scene.json
python3 $P/video-animatic/scripts/timeline_compiler.py --scene film/scene.json --out film/animatic.html
python3 $P/video-animatic/scripts/frame_renderer.py --html film/animatic.html --fps 12 --mp4 film/animatic.mp4
# 4 (previz-to-final lane only, after the animatic is approved) build the upgrade manifest
python3 $P/video-animatic/scripts/shot_manifest.py --scene film/scene.json --out film/shots.json
# 5 Final: switch mode to final, render at full fps, conform
python3 $P/video-animatic/scripts/timeline_compiler.py --scene film/scene.json --mode final --out film/final.html
python3 $P/video-animatic/scripts/frame_renderer.py --html film/final.html --out film/frames
python3 $P/video-conform/scripts/conform.py --render film/frames/render.json --scene film/scene.json --vo film/vo.wav --out film/final.mp4 --run
Exit codes carry the decisions: brief_gate 3 = ask (one question at a time,
recommended answer first), 4 = refuse; beat_board 4 = cut words or raise the
length; motion_linter 4 = fix before rendering; conform 4 = QC failed.
Phase gates
| Phase | Done when | Hand-off |
|---|---|---|
| Brief | brief_gate READY | brief.json with lane, aspect, VO budget |
| Beats | beat_board FITS | scene.json (greybox) |
| Animatic | linter PASS, user watched animatic.mp4 and approved | approved scene |
| Upgrade (optional) | every shot quoted, user said yes to the total | plates in film/plates/, set as plate src |
| Final | conform PASS | final.mp4, captions.srt |
Show the animatic before anything else is polished. Most notes land there, and there they cost seconds.
Shot upgrades with a generation MCP
Only in the previz-to-final lane and only after the animatic is approved.
Use whatever video-generation MCP the user has connected (Higgsfield,
ElevenLabs, Adobe, Canva or another); read its tool list at run time, never
assume model ids. For each shot in shots.json:
- Get the provider's price without generating (Higgsfield
generate_videotakesget_cost: true). Show per-shot and total againstbudget_credits. - Generate only after the user says yes to that total. One batch, no retries on timeout: a timed-out job may still be charged; ask before resubmitting.
- Save the plate, set it as the shot's plate
src, re-render that shot only (frame_renderer.py --shots s2). Text, charts and logos stay as code layers on top, so they never come out garbled.
Data flow: prompts and any reference stills go to that provider under the user's account. Nothing goes anywhere else.
Hard rules
- No render before
brief_gateREADY; no generation without a quote and a yes. - Never generate a real person's likeness without recorded consent.
- Label generated footage as AI-generated where it is published.
- Captions ship with every final (
captions.srt).
Forcing-question library
Ask one at a time, recommended answer first.
- What should the viewer do after watching? Recommended: one action they can take today.
- Where does it run first? Recommended: the platform with the most viewers; other aspects re-render from the same scene.
- How long? Recommended: 15-30 s for social; every extra 10 s needs its own reason.
- Is there a voice-over, and who records it? Recommended: a recorded human read; captions either way.
- May any shot be generated, and what is the credit cap? Recommended: no at first; approve the free animatic, then upgrade one hero shot.
- Do real people appear, and is consent recorded? Recommended: no generated likeness without it.
References
- previz_pipeline.md — why animatic first, what each phase decides
- Sibling skills:
video-beat-board,video-animatic,video-conform