verify - SKILL.md Agent Skill

name: verify description: Verify that a code change actually does what it's supposed to by running the app and observing behavior. Use when asked to verify a PR, confirm a fix works, test a change manually, check that a feature works, or validate local changes before pushing — beyond just unit tests. category: development activation: manual tags: [verify, validate, confirm, qa, smoke-test]

Verify the change

Goal: confirm a specific change does what it claims, by observing real behavior — not by trusting the diff or a passing test alone. $ARGUMENTS may name the change / PR / expected behavior to verify; if empty, infer it from the recent work or the current diff.

1. State the claim precisely

Write down, in one line, what the change is supposed to do and how you'd know it worked ("after the fix, avg([1,2,3]) returns 2, not NaN"). If you can't state a checkable claim, ask for one rather than guessing.

2. Establish the before/after when feasible

If practical and safe, observe the BUGGY behavior first (e.g. on the base revision, or by reverting the change in a scratch copy / cc checkpoint), then the FIXED behavior — so the verification is differential, not just "it ran".
For a pure feature add, exercising the new path is enough.

3. Exercise it for real

Use the run skill's launch patterns to start the app and drive the exact path the change touches (the new flag / route / screen / function) with realistic input — including an edge case the change was meant to handle.
Prefer observing the running app; fall back to a focused script or a targeted test ONLY if the app genuinely can't be launched here (say which and why).
Run the existing tests too if they're relevant, but treat them as supporting evidence, not the verification itself.

4. Verdict

Give a clear verdict with evidence:

VERIFIED — quote the observed output/behavior that proves the claim, and note any edge case you checked.
NOT VERIFIED — show what you observed that contradicts the claim (the bug still reproduces, an error, wrong output).
BLOCKED — state exactly what stopped you (no display, missing service, needs a device/credentials) and the precise command/steps you'd run once unblocked. Never report VERIFIED for something you did not actually observe.