← ClaudeAtlas

double-checklisted

Use before declaring any code change done, when the atlias gate holds a reply for a second pass, or when asked to bug-check, verify, review your own change, or check for everything. Two passes: functional proof, then an adversarial hunt. Also use when tests fail and you need to turn the failure into a fix.
ridelink0/atlias · ★ 1 · AI & Automation · score 72
Install: claude install-skill ridelink0/atlias
# double-check Done means two passes happened and the reply names both. ## Pass 1: functional proof 1. List every file you changed this turn (`harness_progress get` shows them). 2. Run the cheapest real check that would fail if the change were wrong: `node --check`, `JSON.parse`, `py_compile`, the project's own test or sweep script, a request against the running app. `harness_verify {paths}` does the syntax floor. 3. Read the output. A check you did not read did not happen. ## Pass 2: adversarial hunt Re-read each changed file as the person trying to break it. Ask, in order: - Empty, missing or malformed input: what happens on `null`, `""`, `[]`, a corrupt line, a missing file? - Boundaries: first item, last item, exactly the threshold, zero, negative. - Windows: backslashes, drive letters, a BOM at the start of a file, CRLF, a path with spaces, a locked file. - Callers: who assumed the old behaviour, the old name, the old return shape? Grep for them. - Failure paths: which `catch` swallows the error the user would need to see? - Concurrency: two hooks or two sessions writing the same file at once. - Regressions: what worked before that this change could have broken? Run that check too. Fix what you find, re-run pass 1 for the fixed files, and end the reply with one line: `Pass 1: <check> passed. Pass 2: <what was found and fixed, or nothing found after checking X, Y, Z>.` ## Reading a test failure atlias tests and gate messages come in three lines: what happened, wh