Shot-scraper Video Lets AI Agents Record Demo Videos

Simon Willison released shot-scraper 1.10 (announcement; GitHub). The update adds a dedicated video command that records browser sessions from a storyboard file.

The command reads steps from storyboard.yml. It then uses Playwright to capture the full session as video.

New command removes manual steps

Earlier versions of shot-scraper produced screenshots. Users had to stitch frames by hand for any video output. Version 1.10 changes that workflow.

The new command accepts the same YAML format used for screenshots. It passes the sequence to Playwright 1.61.0 screencast support. This produces a clean MP4 without initial blank frames or fixed 800-pixel width. In a hypothetical scenario, a product manager could ask an agent to record a live demo of a CSV-upload feature for an investor update, automatically generating a polished video directly from recent workflow notes without any manual screen recording.

Developers tested the feature on Datasette. The example shows an agent pasting CSV, TSV, or JSON data and creating a new table. The resulting video matches the live session exactly.

Detailed help text supports agents

Simon Willison designed the --help output to be exhaustive. The text includes every flag, required file format, and error condition.

An agent can therefore read the help text and call the command without further human guidance. The goal is to let coding agents produce their own demonstration videos on demand.

Context matters for agent tasks

An agent asked to record a demo needs more than the command itself. It needs the full sequence of actions that produced the state shown in the video. That sequence lives in meeting notes, prior code changes, and project decisions.

remio captures those sources automatically. It stores meeting transcripts, repository activity, and document versions in one personal knowledge base. An agent can query that base before writing the storyboard file - for instance, by asking “List the exact steps discussed in yesterday’s meeting to handle CSV upload in Datasette” - and surface the relevant workflow details. The result is a demo video grounded in actual workflow steps rather than generic examples.

Comparison with existing tools

Most screen-recording utilities such as Loom and OBS Studio still require a human to trigger capture and edit the output. Shot-scraper video removes that manual layer once the storyboard exists.

Other agent frameworks such as Auto-GPT and LangChain can call shell commands, yet they usually lack the surrounding work context needed to build an accurate storyboard in the first place. The combination of detailed command help and accumulated context closes that gap.

Limits that remain

The command works only inside a browser session defined by Playwright. Desktop applications or local tools outside the browser fall outside its scope.

Playwright screencast also records at the browser viewport size. Teams that need multiple resolutions must still run separate passes.

What to watch next

Teams will test whether agents can generate storyboards from meeting transcripts stored in remio. Early results will show up in public agent logs within the next month.

Simon Willison may publish additional examples that combine the video command with other CLI tools. Those releases will indicate how far the pattern of agent-driven documentation can extend.

Broader adoption depends on Playwright updates that improve video quality and reduce frame drops. The next Playwright minor release is the clearest signal to monitor.