"Look at It" Skills: Closing the Verification Loop

Stop asking the human what the screen says

A powerful pattern is a skill that lets Claude observe the effect of its changes instead of asking you to look: a browser skill that loads a page headlessly and reports console errors, failed network requests, the title and optional DOM checks or a screenshot; a game or app runner that starts the program and returns the log; a test-and-report skill that runs the suite and summarises failures; an API smoke test that calls an endpoint and prints status and a trimmed body. This closes the loop edit, run, observe, fix that otherwise needs the human to describe what happened. Design such skills so that they print concise, structured findings (not megabytes of logs), have a timeout, clean up processes they start, and make clear what was and was not verified. The description should say when to use it ("use to verify your own web changes").

The verification loop

Each step produces something Claude can read.

edit code
  -> run (start app / load page / run tests)       <- a skill script, with a timeout
  -> observe: console errors? failed requests? failing test? status code?
  -> concise structured findings back to Claude   <- JSON or short labelled lines
  -> fix and repeat until the findings are clean
  -> final report states exactly what was verified (and what was not)

Put a timeout on every run

An app or browser that hangs should not freeze the whole session.

Quick check: What does a "look at it" skill remove from the workflow?

• The need for a computer
• The need for code
• The need for tests
• Asking the human to describe what is on the screen or in the log