← ClaudeAtlas

verify-worklisted

Use before claiming any task is done, complete, or fixed. A final verification pass that catches unverified claims, unintended changes, and leftover debris before reporting to the user.
suleymanbyzt/agent-helm · ★ 1 · AI & Automation · score 72
Install: claude install-skill suleymanbyzt/agent-helm
# Verify Work Goal: never tell the user something works without having watched it work. Run this before every "Done" report. "It should work" is not a verification. ## Checklist 1. **Run the tests.** The relevant test suite actually executed, actually green — in this session, after the last change. A test run from before your final edit proves nothing. 2. **Run the thing.** Where possible, exercise the changed behavior for real: call the endpoint, run the command, hit the code path. One real execution beats ten confident explanations. 3. **Read the full diff.** - Every changed line is intentional and belongs to this task. - No accidental behavior changes, no unrelated "improvements". - No leftover debug output, commented-out code, or TODO placeholders. 4. **Check the tests themselves.** No test was weakened, skipped, or deleted to get to green. If one was — with approval — it is mentioned in the report. 5. **Check the claims.** Everything you are about to tell the user is something you observed, not something you infer. Anything unverified is reported as unverified: "not tested: X" is honest; silence is not. 6. **Second opinion (L2/L3).** If an advisor/reviewer model is available, have it review the result before you present it. If not, re-read the diff once more with fresh eyes on the riskiest part. ## Then report ``` Done: <one sentence> Verified: <what you ran and what you observed> Recorded: <journal file p