verifying-before-done — Claude Code skill

What verifying-before-done does

The trigger is the sentence the agent is about to write: shipped, fixed, all green. This skill makes it earn that sentence. Instead of judging by the diff, the agent exercises the path where the change meets its caller — sets the switched flag, sends the handler its request, hits the error branch — and reports what came back.

A clean suite is treated as suspicious, not reassuring. The skill asks whether the test actually ran and whether it can go red at all; a check that never fires yields a confident and wrong “done”. It also keeps three levels apart: present in the repo, running live, and confirmed in the outside system. Evidence from one never stands in for another.

Before wrapping up, the agent sweeps the diff for stray edits, missing docs and acceptance criteria that were only roughly met. Changes on declared critical paths, plus irreversible migrations, need a reviewer who did not write the code. The skill requires principal-engineering as background, so that one goes in first. Destructive operations are not run live: with no safe target, verification happens around them and the untested path gets named.