video-kinetic

Kinetic video

Turns a short script into a punchy motion-graphics video: every word pops in at the moment it is spoken, in a heavy display font; key words glow and punch the camera; between beats the screen is cut by a glowing slash, shatters, whips or zooms through; optional 3D-tilted cards and node flows show what the user's script lists. It is rendered to MP4 by video-core's Remotion engine on the user's computer (Claude desktop app, Code tab, Local; Windows, Linux or macOS). <skill> is this folder; <video-core>, <video-voice> and <video-qa> are sibling skills. The look is original: nothing of any other creator's video is used.

Use when

• The user wants a hype, launch, explainer or "motion graphics" video where bold words appear as someone speaks, with fast stylish transitions.
• They give a topic or a script. If only a topic, write a short script of their facts (references/script.md); never invent a claim, number or quote.

Don't use when

• The user wants their own footage cut up, a product-photo edit or generated people. This library does not make those.
• The words would need facts the user has not given. Ask for them instead.

Steps

1. Write projects/<slug>/script.md: 5 to 12 short lines, one idea each (references/script.md). Optional attributes under a line: emph: (key words), layout: slab (one giant word on a colour block), transition:, exit:, card: Title|row=tag;row and flow: a>b>c (only content the user supplied).
2. Pick a preset (volt lime, ember orange, ice cyan, grape violet; --accent #rrggbb overrides). Run node <skill>/scripts/plan.mjs --project projects/<slug> --preset ember.
3. Voice and timing: <video-voice>/scripts/voice.mjs (the local synthetic voice, or recorded.mjs <project> --file mic.wav --speaker self for the user's own recording, which also writes the word times), then align.mjs (real word times from whisper.cpp), then <video-core>/scripts/timing.mjs.
4. <video-core>/scripts/check.mjs, then render.mjs --draft. The render adds the sound itself: a whoosh on each cut, an impact on slabs, pops on the beats and a music bed under the voice.
5. Run node <skill>/scripts/sync.mjs projects/<slug> (--allow-estimated without whisper.cpp) and <video-qa>/scripts/qa.mjs, open qa/contact-sheet.png and judge it with the checklist in references/craft.md. Fix by editing script.md attributes or video.json and re-planning (a new folder) or editing in place; at most 2 revisions, then show the user.
6. Render the final (no --draft), run QA again, show the MP4 path and contact sheet, offer the publish pack (video-publish). Never upload.

Budget

• Search first (grep, glob, symbol search). Read line ranges, not whole files.

• Read at most 8 files per step. If you need more, write down what you learned and narrow the question.

• Do not re-read a file you just wrote or edited.

• Prefer the smallest command output that answers the question: head, tail, quiet flags, filtered lists.

• Run each script once per revision. Do not read words.json or the render folder.

Output

• The project folder, the preset, the number of beats and transitions, the sync result and the QA result.

Done when

• sync.mjs and video-qa pass, the contact sheet passes references/craft.md, and every word on screen came from the user's script.
• Nothing was uploaded, and a recording of someone's voice was used only with their confirmation.