shot scraper 1.10's video subcommand drives Playwright (the browser automation engine) from a small YAML storyboard to record a coding agent's work in a real browser. The result is easier to review than a patch, but it is still the agent grading itself.
Run a coding agent against a feature. When it finishes, instead of handing back a patch and a wall of log output, it can now leave you a short screen recording of the new feature actually running in a browser. That is the user-visible promise of [shot-scraper video](https://simonwillison.net/2026/Jun/30/shot-scraper-video/), a new subcommand shipped in shot-scraper 1.10 on June 30, 2026.
shot-scraper is an established command-line tool for capturing screenshots and scraping web pages with Playwright, the browser-automation engine. The new video subcommand is additive rather than a from-scratch framework. You feed it a storyboard.yml describing a routine, and Playwright records the browser session as it walks through that routine against your app.
A storyboard is a short YAML file naming the output file, the URL to open, the viewport (typically 1280x720), whether to show the cursor, and a sequence of scenes. Each scene is a list of steps: pause, click a selector, fill an input, wait for text or a URL. A trimmed excerpt from the reference storyboard looks like this:
``yaml
output: /tmp/datasette-bulk-insert-demo.webm
url: http://127.0.0.1:6419/demo/tasks
viewport: { width: 1280, height: 720 }
cursor: true
scenes:
do:
`
The command can also start a local server before recording, pass a JSON cookie file via --auth for logged-in flows, and emit MP4 or WebM. The reference demo runs against Datasette, an open-source tool for publishing data, exercising an in-development feature that creates tables from pasted CSV, TSV, or JSON.
The limit: this records what the agent did, not whether the agent was right. A 12-second .webm is easier to review than a 400-line patch, but it can also hide failure modes a test suite would surface. A self-recorded success is weaker evidence than a failing test the agent also recorded.
Willison's framing is narrower than the usual hype. He has argued for years that coding agents should ship demos of their work, not just patches; shot-scraper video is the latest step in that line, alongside his earlier showboat and rodney tools. He also notes that both the example storyboard and the underlying Datasette feature were written end-to-end by GPT-5.5 at its xhigh reasoning setting inside Codex Desktop, per his own writeup, a provenance worth flagging as self-reported rather than independently verified.
The plumbing took longer than the AI part. Willison filed the original issue in February 2024 and only shipped this week, because Playwright's recording primitives were not stable enough until Playwright 1.59's screencast API and a recent fix removed the 800-pixel width ceiling.
What to watch next: whether other agent harnesses adopt a storyboard-shaped demo artifact as a default output, and whether shot-scraper video` starts showing up in agent-recorded pull requests as a portable, human-reviewable trail.