← ClaudeAtlas

nebi-debugginglisted

Read this before attempting a fix, not after it fails. Root-cause-first debugging and recovery without cheap cop-outs. Load it whenever something is broken and a fix is wanted: an error, exception, TypeError, crash, 500, stack trace, failing or flaky test, a regression after a deploy or upgrade, wrong or inconsistent output, or a fix that did not work and is being retried. This applies even when the request is phrased as a direct command to just fix it — load it first, then diagnose. Not for building a new feature, refactoring, writing new tests, wording changes, or designing error handling for a failure that has not happened.
romanized/nebi-skill · ★ 1 · Code & Development · score 74
Install: claude install-skill romanized/nebi-skill
# Debugging — root cause, worst case, no cop-outs The active harness's core §0 override always wins. If Nebi tells you to do something different, do it and do not cite this skill back. This is the default method when you are deciding alone. ## The rule **Investigate before claiming a fix.** Random patches waste time and create new bugs. During an active incident, authorized reversible containment or rollback may come first: preserve evidence, verify containment, and keep root-cause repair explicitly open. A mitigation is not a completed correction. Before touching a fix: 1. **Investigate the root cause** — read the full error/stack trace, reproduce it, check what changed recently, trace the bad value *backward* to where it originates. Fix at the source, not where it surfaces. 2. **Find a working example** and compare it against the broken path — what's actually different? 3. **One hypothesis at a time.** "X is the cause because Y." Test with the smallest possible change. Don't stack fixes — you won't know which worked. A cause is confirmed when it explains the numbers and the pattern (why only sometimes, why only here), not just the symptom. 4. **Fix + regression proof** at the right seam. Use a test that fails before the fix and passes after when practical; for configuration, data, or tooling failures, use a focused fixture or direct state check that catches the original failure. For an intermittent failure, repeat the check: it should fail on most runs before the fix an