A video for every release

A clip for every release, from the same cut.

Updated November 2026

How it works

Three steps, in the order they happen.

video/actions.jsonin the repo
{ "url": "http://localhost:5199/",
"steps": [
{ "id": "tool-rect", "do": "click" },
{ "id": "box-landing", "do": "drag" },
{ "do": "type", "text": "Landing" } ] }

Script the feature’s flow Write an actions.json for the new feature: wait, hover, click, type, scroll, move. Keep it in the repo beside the feature; it is the replay recipe.

Claude Code
$ vos record --actions actions.json
✓ 44 s take · 37 steps
$ vos push take --label "v2.1"

Record in CI or by agent The vos CLI records it headless with machine-readable output and exit codes, and the GitHub Action runs it on the pull request and posts the video there.

vos.so/vos/87d517ddplays the latest
v2The same cut, pushed again
v1First version

Keep it at one link Push it to vos.so and the release gets a link that plays. Next release, run the same script: the link plays the new version and keeps the old one.

Why it is worth doing

Release notes with a moving picture get read, and nobody staffs a video editor per release. With a scripted recorder the video becomes part of the release: the flow lives next to the feature, the take renders in CI, and the agent that shipped the code ships the clip. When the UI moves, the same script records again and the cut follows.

The tools that do this, compared →

What stays in the repo

Three small files per video, kept beside the code they show. actions.json is the click path the recorder drives, and it is the next release’s script too. doc.json is the cut as it was last signed off: the zooms, the trims, the captions, the music. vos.json says which video on vos.so each push adds a version to.

Nothing else is needed to make the video again. The footage is not kept in the repo, because the script makes it.

What a release runs

The recorder writes down when every step ran. So the second command can re-time the old cut onto the new footage by step: a zoom placed on the checkout button follows the checkout button, even when a new banner moved everything down.

With --strict, a selector that no longer matches fails the job, so a release never ships a video of the wrong screen.

vos record --actions media/actions.json --out take --strict
vos plan take --reuse --from media/doc.json
vos push take --label "v2.1"

The whole loop, with the CI workflow →

What does not carry over

A step that moved, or a control that left the interface, cannot follow the old cut. It is named in the report, never guessed at, and that list is the release’s real to-do: usually one or two edits, not a new edit.

A flow behind a sign-in needs a session where it is recorded. In CI that means the signed-in state your tests already produce; a session you opened by hand stays on your machine.

Recording behind a login →

vosso does this for itself

The clips on our what’s new page are made this way. A release’s script lives in our repository, and a workflow records the public ones again, on the live product, on every release tag, and on any pull request that changes a script, where it leaves one comment with the video’s page.

Since October 2026 a test fails our build when a release people can see ships without a clip, unless the entry says why it has none.

What’s new, with its clips →

Questions

How do I automatically create videos for release notes?

Keep a scripted browser flow (actions.json) next to each feature and run the vos CLI or the GitHub Action in your release pipeline: record the take headless, then render it. Edits are patches to a JSON document rather than re-recordings, so the output is stable enough to review.

How do I keep a product demo up to date when the UI changes?

Keep the click script in the repo and record it again on each release. The recorder notes when every step ran, so vos plan --reuse re-times the existing cut onto the new footage step by step, and names whatever could not follow: a step that moved, a control that is gone. You fix those and nothing else.

Can this run in GitHub Actions?

Yes. The whole loop is headless. The vosjs/action Action records the script, carries the committed cut onto the new footage, pushes a version labelled for the release, and keeps one comment on the pull request with the video’s page. It needs a content key as a secret; without one nothing is pushed and the comment says so.

Can an AI agent make my changelog videos?

Yes. The pipeline is built for agents: no TTY needed, machine-readable output, and the take’s doc.json is the whole edit surface. An agent that shipped a feature can write the flow, record it, tune the zooms and hand you a link to the video.

Something else? hello@vos.so

Ship the video with the release.