Agent-Chorus 能力展示叙事短片

Agent-Chorus Capability Reel Story Film

Turn one product request into a brisk, four-surface story that shows the work and ends on the brand.

Style: Four-Surface Chorus. Precise and propulsive, with each repeated prompt revealing a new piece of visible progress. Follow references/style-guide.md exactly.

Use this skill when

• A founder introducing a workflow product
• An app developer showing a new capability
• A marketing manager preparing a social launch

Do not use

• Documentary footage, character animation, or photorealistic product demonstrations
• Campaigns that require unsourced growth claims or fabricated customer quotes
• Detailed tutorials that need lengthy narration or sustained screen recordings
• To reproduce another company's brand, logo, product UI or copyrighted characters.

Inputs

• Product name, one-sentence purpose, and intended audience
• One real user request or launch scenario to anchor the story
• Screenshots, interface copy, or verified descriptions of up to four product surfaces
• A verifiable outcome or artifact to show as proof
• Preferred call to action and destination

If something essential is missing, ask one short question before starting. Never invent facts, numbers, quotes or customer names; mark gaps as [needs source].

Workflow

1. Read references/motion-contract.md and references/style-guide.md. Their rules are fixed for this skill.
2. Fill references/scene-map.md from the brief (4-8 scenes, 20-45 s). Never invent product numbers, names or quotes; ask for anything required that is missing.
3. Copy references/composition.html to film/composition.html. Keep the stage, the ambient layer, the progress line and the fit script.
4. Delete the example scenes you do not use, set each scene's text, data-start, data-duration and the absolute animation-delay values, and set <html data-duration> to the sum of scene durations.
5. Apply the style tokens in the :root block exactly as given in the style guide. Do not add colors or webfonts.
6. Open film/composition.html in a browser to preview; it plays in real time. Reload to replay.
7. Render: node scripts/render.mjs film/composition.html --out film.mp4 --fps 30 (install once: npm i -D playwright-core && npx playwright install chromium; FFmpeg must be on PATH).
8. Run the validator below and fix every problem it reports.
9. Extract stills around each cut (ffmpeg -ss <t> -i film.mp4 -frames:v 1 cut.png) and check that nothing is cropped, overlapping or frozen.
10. Run the validator and fix every problem it reports:

``bash node scripts/validate-video.mjs film.mp4 --min-seconds 20 --max-seconds 45 ``

What makes this excellent

• Write the story as hook, friction, four repeated product beats, proof, and end card; keep the total between 20 and 45 seconds.
• Turn the user's scenario into one prompt of no more than five words, then repeat it verbatim across the four surface beats.
• Assign capture, plan, build, and publish distinct interface states while retaining a fixed prompt position and consistent framing.
• Make each hard cut land when the next surface answers the previous surface's unresolved question.
• Use supplied screenshots or verified interface copy for product claims; label invented interface details as illustrative in the production notes.
• Hold the proof scene long enough to read its artifact, and omit any metric that lacks a source.
• End with the product name and one specific action, leaving the final card visible for at least two seconds.

Output contract

• film.mp4: H.264, 30 fps, 1280x720, 20-45 seconds.
• film/composition.html: the editable single-file composition that produced it (re-render any time with scripts/render.mjs).
• A hand-off note: the filled scene map, total duration, anything left as a placeholder, and the validator output.
• Deliver a locally rendered wide MP4 and the editable HTML/CSS composition.
• Include a scene-by-scene timing and copy sheet with the total runtime.
• List which on-screen claims use supplied evidence and which interface states are illustrative.

Reference demo and agent run

The preview is a reference render of the bundled composition with sample copy. In a real run the agent rewrites the scenes from the user's brief and assets, re-times the composition and renders a fresh MP4 that meets the output contract above.

Failure rules

• Never hand over a render that fails the validator.
• If the user's screenshots are low resolution, rebuild the surface in HTML instead of upscaling them.
• Music and footage must be supplied and licensed by the user; never download assets.