vos skills

Agent skills for vos, the open video engine and the vos CLI. Install once and your coding agent records the real product, cuts it as data, and ships the release's media from one take: the demo, the store listing, the Product Hunt gallery, the social cuts, the OG card, kept current with every release. It can also author and remix vos animation programs.

npx skills add vosjs/skills

Works with any agent that reads the Agent Skills format (Claude Code, Cursor, Codex, and friends). Claude Code users can also add the repo as a plugin marketplace: /plugin marketplace add vosjs/skills.

The catalog

Skill	What it does	Needs
product-video	Record and produce a shippable product demo video of a website, app, or feature: the agent drives the browser from a JSON script, the recording auto-plans zooms, and every edit is a JSON patch on doc.json. One take also yields exact-size store screenshots, tiles, and social stills.	npm i -D @vosjs/cli, a Chromium, network
vos-authoring	Author vos animation programs (VosConfigJson) from scratch: schema, dialect rules, params/presets knobs, local validation with @vosjs/core.	npm i @vosjs/core (validation), @vosjs/cli (render)
vos-remix	Remix a program from the vos.so gallery: fetch its config, edit knobs or code, validate locally, push a private copy back with lineage — then iterate it as versions. Ships the params/Looks doctrine (2–4 intent-named knobs, every knob verified to act).	npm i -D @vosjs/cli, a content key or remix grant
vos-create	Create vos programs in the user's own style: pull their vos.so folder, read every recipe (.md) and exemplar in it, find the declared seed, author work that follows them, check and render and judge locally, and push into that folder as attributed private versions. Recipes carry the style; the user's ask decides what gets made.	npm i -D @vosjs/cli, a content key, a folder on vos.so
vos-cut	Cut a screen recording somebody made into a product video, from its evidence: vos digest lists the moments the cursor track says mattered with a footage frame and a crop each, the agent narrates them into beats, edits doc.json by exception over the planners, verifies with stills, pushes an attributed version, and iterates with the human through the version chain. A signed-off take becomes a series' seed: its folder's CUT.md carries the judgment, vos plan --style carries the numbers.	npm i -D @vosjs/cli, a Chromium, a content key for hosted takes
launch-kit	One take → the full launch asset kit: YouTube demo + thumbnail, X/LinkedIn cuts, Chrome Web Store screenshots + tiles, Product Hunt gallery, GitHub README loop, OG cards. Per-channel specs ship as data and every asset is verified against them in a kit.json manifest.	npm i -D @vosjs/cli, ffmpeg, a finished take
vos-migrate	Turn an existing Loom, Screen Studio, Cap, or plain mp4/webm screen demo into an editable vosso take, so the next version is a re-render instead of a re-recording.	npm i -D @vosjs/cli, the file
vos-footage	Hand a Remotion, HyperFrames, or any video composition a clean full-bleed clip of the real product, recorded and auto-zoomed with vos, always backed by an editable take on the maker's shelf.	npm i -D @vosjs/cli, a Chromium, a content key

One CLI, one package: @vosjs/cli is the MIT vos binary with the engine verbs (local, no account), the take pipeline (record, plan, frames, render on screen recordings) and the vos.so platform verbs (fetch, push, pull, login). Everything that runs on your machine is open source, and vos.so, the hosted platform, is the product. (The plugin was previously published as @vosso/cli; that name keeps working as a forwarding shim.)

Why these are different

• Real footage, witnessed. product-video drives a real browser over a real page. The output is a recording of software actually working, not a synthesized imitation of it.
• Edits are data. A take's every decision (zoom spans, trims, speed, tilt, audio) lives in doc.json. Agents patch JSON and re-render; nothing re-runs the browser.
• Deterministic renders. Same take, same doc, same frames, on any machine. Video work becomes reviewable the way code diffs are.

Quality

Every skill ships with eval scenarios in evals/, run before each release on two model tiers; a launch-kit run ends on one deterministic number, the kit verifier's problem count (evals/count.mjs), published on the release's changelog entry. The product-video skill carries the taste rubric its outputs are judged against; if a run ships something ugly, that failure becomes a new eval scenario.

Links

• Engine: github.com/vosjs/vos · vos.so/engine
• Agent docs: vos.so/llms.txt · vos.so/agents
• The gallery (every video is a program): vos.so/gallery

Contributing

Issues and PRs welcome. Keep SKILL.md under 500 lines, references one level deep, and run node scripts/check.mjs before pushing. Skill text must match what the released CLI actually does; a claim the CLI can't honor is a bug.