Building verification loops so Claude checks its own work

Key points

• Claude Code runs a four-stage loop on every prompt: gather context, take action, verify the work, then respond. 0:00
• Passing tests, type checks and linters does not prove a change does what was intended; the video says a manual check is what proves it works. 0:24
• Codify the manual checks in the project (for example opening the page and reading the console, calling an endpoint, or tapping through simulator screens) so Claude can run them itself. 0:37
• Start from the verify skill that ships with Claude Code, then extend the generated skill with the checks your project needs, such as a performance trace for layout shift. 1:01
• A good verify skill states when Claude runs it, what to do when a check fails, and what proves each check passed, and the more measurable a check is, the easier it is for Claude to tell if it passed. 1:37

Small clues in the verify demo

• Performance trace result — Text reports CLS went from 0.19 to 0.00 and that a shipping banner caused the layout shift.
• Claude Code edit log — Claude's log lists edits to banner.js and styles.css, with a prompt to fix the layout shift.

Questions to check yourself