What Actually Counts as 'A Check Claude Can Run'
You want to give Claude Code a genuine verification target, instead of letting it decide the work "looks done."
Which of these actually counts as a check, in the sense this subject means?
The correct answer is "Something with a clear pass/fail signal, like a test suite."
A real check is anything that returns a concrete pass/fail result Claude can read and react to: a test suite, a build's exit code, a linter, or a screenshot diff against a design. Claude re-reading its own diff, or rating its own confidence, is still just Claude's own judgment of whether the work looks done — the exact failure mode a check is meant to remove. And waiting for a user to notice a problem in production is not verification at all; it's the absence of it.
Share this question