← ClaudeAtlas

debuglisted

Systematic root-cause debugging — hypothesis-driven analysis, evidence gathering, and minimal reproductions instead of guessing. Use whenever the user reports a bug, error, unexpected behavior, test failure, crash, an exception or thrown error, or performance issue, or says something "doesn't work", "broke", "stopped working", "regression", "flaky". Also triggers on pasted stack traces, error messages, log snippets, or when the user asks "why is this happening?", "why does this throw?", "what's wrong?", "marvin debug this", or "help me debug" — including "walk me through why this throws" when the user wants the cause found or the failure fixed. (If they only want to understand how existing logic behaves and are explicitly not asking for a fix, that is explain, not debug.) Applies to any language, framework, or runtime.
real-case/marvin-toolkit · ★ 0 · Code & Development · score 68
Install: claude install-skill real-case/marvin-toolkit
# Debug Find the *actual* root cause before touching any code — most debugging time is wasted on premature fixes. ## Core principle **Read, don't guess.** Gather evidence → form hypotheses → verify the most likely one → only then fix. Jumping to a fix on a hunch is the #1 debugging anti-pattern. ## This skill is the door to `marvin-debugger` The hypothesis-driven methodology lives in one place — the **`marvin-debugger`** agent — so the debug skill and `/marvin:task-start` Step 3B share a single source instead of duplicating it. This skill is the **interactive** path: you stay in the conversation and apply the fix; the agent does the isolated, evidence-first analysis. ### 1. Dispatch the agent If Task-tool is available, invoke **`marvin-debugger`** with everything you know about the bug — the symptom, any stack trace or error text, repro steps, and the suspected area. It runs read-mostly (it may write a throwaway reproducer) and returns a structured report: **Symptom · Reproduction · Evidence · Hypotheses · Root Cause · Fix Approach · Regression Test · Siblings · Lesson**. ### 2. Act on the report - **Root cause confirmed** → apply the agent's **minimal** fix approach with the user. Write the **regression test first** (it must fail on the current code), then apply the fix and confirm the test goes green and the original reproducer is gone. Then grep for the **siblings** the agent listed and decide with the user whether they're in scope. - **Unconfirmed** → the agent r