vibe-scene
Skill185 starsChanged todayMITView on GitHub
Author, repair, render, and inspect VibeFrame scene projects built from STORYBOARD.md and DESIGN.md.
- Reads credentials
In this repository11 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- releaseSkill
- sync-checkSkill
- testSkill
- vibe-pipelineSkill
- vibe-sceneSkill
- releaseSkill
- sync-checkSkill
- testSkill
- vibe-pipelineSkill
The file
We split this skill into the rules it gives an agent
12 rules. About 0 are common, 4 are distinctive, and 8 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
8
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 18
Use this skill when the user wants editable HTML-based scene composition, storyboard-to-MP4 work, or a host-agent authoring loop.
- Distinctive: line 21
Use the top-level commands for the main project lifecycle:
- Distinctive: line 40
Default recommendation for public demos and reproducible dogfood runs:
- Very specific: line 54
Use --skip-backdrop when you want a low-cost composition test.
- Very specific: line 60
Use this when Claude Code or another coding agent should be the reasoner that authors compositions/scene-*.html.
- Very specific: line 72
If vibe build returns a needs-author plan:
- Distinctive: line 74
Read the compose plan.
- Very specific: line 82
Every timed element needs data-start, data-duration, and data-track-index.
- Very specific: line 85
Do not use Date.now(), Math.random(), or network fetches in render paths.
- Very specific: line 101
Run vibe schema scene.<subcommand> before using less common flags.
- Very specific: line 155
Never author scene HTML before DESIGN.md exists.
- Very specific: line 190
Run lint with --json --fix.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
<!-- GENERATED by `pnpm agent-sync` from .agents/skills/vibe-scene/SKILL.md — DO NOT EDIT. -->Vibe Scene
VibeFrame scene projects are editable video projects built from:
STORYBOARD.md- beats, narration, backdrop cues, duration hintsDESIGN.md- visual identity, palette, typography, motion rulescompositions/scene-*.html- per-beat HTML compositionsindex.html- root timeline that references scene compositionsassets/andrenders/- generated inputs and final outputs
Use this skill when the user wants editable HTML-based scene composition, storyboard-to-MP4 work, or a host-agent authoring loop.
Use the top-level commands for the main project lifecycle:
vibe init my-video --profile agent --ratio 16:9
vibe build my-video
vibe render my-video -o renders/final.mp4The vibe scene ... namespace is the lower-level authoring surface:
install-skill, compose-prompts, list-styles, add, and lint.
Pick The Right Path
| Path | Use when | Commands |
| Host-agent authoring | Claude Code/Codex/Cursor writes the scene HTML from a plan (default) | vibe build, then vibe scene compose-prompts |
| Deterministic template | CI or a human demo should produce scene HTML without a coding agent (no LLM call) | vibe build --composer template |
| Single-scene draft | You need a quick template scene or fallback HTML | vibe scene add |
Default recommendation for public demos and reproducible dogfood runs:
vibe init my-video --profile agent --ratio 16:9
# edit DESIGN.md and STORYBOARD.md
vibe build my-video \
--tts kokoro \
--skip-backdrop \
--skip-render
# (the host agent authors compositions/scene-*.html, then re-run vibe build)
vibe scene lint index.html --project my-video --fix
vibe render my-video -o renders/final.mp4 --quality standardUse --skip-backdrop when you want a low-cost composition test. Remove it when
the demo should exercise OpenAI image generation from each beat's backdrop
cue.
Host-Agent Loop
Use this when Claude Code or another coding agent should be the reasoner that
authors compositions/scene-*.html.
vibe build my-video \
--tts kokoro \
--skip-backdrop \
--skip-render
vibe scene compose-prompts my-video --jsonIf vibe build returns a needs-author plan:
- Read the compose plan.
- Author each missing
compositions/scene-<id>.html. - Run
vibe scene lint index.html --project my-video --fix. - Fix remaining lint errors by editing the HTML.
- Run
vibe render my-video.
Hard rules for authored scene HTML:
- Every timed element needs
data-start,data-duration, anddata-track-index. - Timed elements must have
class="clip". - GSAP timelines must be paused and registered on
window.__timelines. - Do not use
Date.now(),Math.random(), or network fetches in render paths. - Route final audio through root-level
<audio>elements when possible. - For factual, typography-heavy videos, keep claims and layout in HTML/CSS/JS; treat provider assets as inputs rather than the whole product.
Low-Level Scene Commands
vibe scene install-skill my-video --host auto
vibe scene compose-prompts my-video --json
vibe scene list-styles
vibe scene add intro --project my-video --style announcement --headline "Hello"
vibe scene add badge --project my-video --style simple --lottie assets/logo.lottie --lottie-position bottom-left
vibe scene lint index.html --project my-video --json --fixRun vibe schema scene.<subcommand> before using less common flags.
Overlay a Lottie animation (.json/.lottie) on a scene with --lottie <path>,
tuned by --lottie-position (full|center|top-left|top-right|bottom-left|bottom-right),
--lottie-scale, --lottie-opacity, and --lottie-no-loop. The scene timeline
drives it frame by frame from a runtime copied into assets/vendor/dotlottie/
(no provider call, free).
Takes: Keeping A Generation You Liked
vibe build records every beat asset it generates (narration, backdrop,
keyframe, video, music) as a take, and keeps each take's bytes. When a
regeneration came out worse, put the earlier one back instead of paying again:
vibe take list --beat <id> --json # inSlot marks what the slot holds now
vibe take promote <take id> --dry-run --json
vibe take promote <take id>Later builds keep a promoted slot until that beat's cue is edited or
vibe build --force runs. vibe take log --json totals what the project
spent, split into provider billing and estimates. vibe take prune --dry-run
shows media copies safe to remove (--slots-only frees more, but those takes
can no longer be promoted from a copy; ask first). See docs/take-graph.md.
A beat video that is close but not right can be continued instead of
regenerated: vibe take extend <take id> adds time after its last frame,
vibe take edit <take id> --prompt "<the change>" revises it (Grok and
Gemini Omni; Kling legacy models extend only). Both are paid: run
--dry-run --json first, confirm the spend with the user, and pass
--max-cost. The result is a new take for the same slot; promote it to use it.
Review reports link findings to takes: an issue about one beat lists the
takes filling its slots in issues[].takes, and a finding about generated
content adds take edit / take promote actions to nextActions. Use those
rather than regenerating the whole beat.
Narration Word-Sync
During vibe build, generated narration is transcribed at the word level
(ElevenLabs Scribe when ELEVENLABS_API_KEY is set, else OpenAI Whisper) to
assets/transcript-<beat>.json whenever narration exists and one of those keys
is configured. Those word timings are passed into each beat's compose
prompt so captions / kinetic typography can be synced to speech.
- On by default; pass
--skip-transcriptto opt out (no transcription cost). - Neither key → silently skipped; narration still plays, just without word-level animation timing.
- Timings are advisory and visual-only — the composer never adds audio/SFX from them. For long narration the prompt downgrades to coarse phrase-level anchors.
STYLE And STORYBOARD Guidance
Never author scene HTML before DESIGN.md exists. Treat it as the hard gate
for visual decisions:
- palette and contrast
- typography and hierarchy
- layout density
- animation timing and transition style
- what to avoid
Scenes live in scenes/, one markdown file per scene, playing in filename
order. Cues go in the frontmatter, direction in the body:
---
type: Scene
narration: "The first sentence the viewer hears."
backdrop: "Specific visual prompt for the scene backdrop"
duration: 5
---
What the scene should show.STORYBOARD.md holds only project frontmatter and direction prose. Beat
headings left in it are ignored and reported - move them into scenes/, or
run vibe storyboard migrate on an older single-file project.
Lint Feedback Loop
vibe scene lint index.html --project my-video --json --fixRecommended loop:
- Run lint with
--json --fix. - Fix any
errorfindings in the referenced scene file. - Re-run lint.
- Cap retries at 3; if errors persist, replace the scene with a simpler
vibe scene add ... --style simple --forcedraft and explain the tradeoff.
Warnings can be acceptable for demos only when the final render is visually correct and deterministic.
Render Checklist
vibe doctorreports Chrome and FFmpeg as available.vibe scene lintexits 0 or only acceptable warnings remain.index.htmlreferences everycompositions/scene-*.htmlfile needed.vibe render my-video -o renders/final.mp4writes the expected MP4.- Final
media infoorffprobeconfirms the intended duration, fps, and codec shape.
Common Failures
Root composition not found: runvibe buildonce so the render scaffold is created, or usevibe init --profile full.needs-author: expected when the host agent authors scenes (the default); author the listed scene files and rerun build.OPENAI_API_KEY not set: use--skip-backdropfor a local composition test, or configure the key before generating backdrops.- Render hangs at 0%: run
vibe doctor; install Chrome or setCHROME_PATH.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

