shot-scraper video: Your AI Coding Agent Can Now Record Its Own Work

The criticism that haunts agentic coding tools — “how do I know what it actually did?” — now has an automated answer. shot-scraper 1.10 gives AI agents the ability to record video demonstrations of their own work.

What Happened

Simon Willison released shot-scraper 1.10, adding a shot-scraper video command that accepts a storyboard.yml configuration file and uses Playwright to record a complete video of a web application workflow. The YAML format specifies server startup, target URLs, viewport size, and granular user actions — clicks, form fills, delays — making the full demo reproducible and machine-readable.

Willison designed this explicitly for coding agents. As he describes it, an agent that implements a feature should be able to “produce demos of its work” — not just commit the code and call it done. The storyboard format uses Pydantic for validation, which means agents can generate valid storyboard files reliably without guessing at the schema.

Why It Matters

Agentic coding tools face a trust problem that tests don’t fully solve. A test suite verifies that a function returns the right value; it can’t show a reviewer what the user experience actually looks like after the agent reworked a UI, rewired navigation, or added new form states. Video closes that gap.

The deeper shift is documentation as a first-class agent output. Most coding agents today produce code and a commit message. A storyboard-driven video adds a third artifact: a verifiable, human-reviewable record of the agent’s functional intent. That matters for code review, handoffs, and audit trails in regulated environments where “the agent changed it” isn’t sufficient accountability.

The format also enables replay. A storyboard generated during implementation can be re-run against future versions to verify nothing regressed — creating a visual regression test at the moment of implementation, without a separate authoring step.

Source: Simon Willison's Weblog

Content created with AI assistance and reviewed for accuracy.