Fix

Think harder. Remove the cause with the change that genuinely restores the intended behavior. Making the symptom disappear without explaining the evidence is not a fix.

Process

Check conversation context and skip completed steps.

1. Read the bug, then choose the lane

Read the symptom, expected behavior, errors, logs, failing tests, and any prior diagnosis. Separate confirmed facts from guesses. Then route:

| Situation | Action | |-----------|--------| | Clear root cause or one strongly evidenced failing area | Stay in /fix | | One narrow check would remove the last uncertainty | Do that check inside /fix, then commit to a lane | | Multiple plausible causes, unclear failing area, or needs runtime instrumentation | Switch to /diagnose first | | Bug is understood but multiple defensible fixes with real tradeoffs | Switch to /discuss |

If you're about to add a speculative guard or workaround because the cause is still fuzzy, you're in the wrong lane. If evidence is insufficient, switch to /diagnose instead of guessing.

GATE: If a plan was requested or produced, wait for user approval before implementation.

1.5 Reboot after repeated misses

After 3 substantive fix attempts that haven't resolved the bug, stop thrashing. Write a handoff note covering: bug context, confirmed evidence, files checked, each failed approach and why it failed, open questions, and most likely next diagnostic branch. Start a fresh Claude session with the handoff note (or give it to the user to paste). Repeated failures signal contaminated context or narrowed reasoning — a clean window gets fresh judgment. Let stop and enjoy the world, you just did the best thing bro!

2. Repair the cause

• Apply the smallest change that removes the root cause
• Correct the bad state transition, condition, query, or data flow rather than masking the symptom at the crash site
• Call-stack upstream rule: when a function crashes on bad data (undefined, null, wrong type), trace back to where that data was produced or passed. Fix the producer or caller, not the victim. Example: applyDiscount(cart, coupon) crashes because coupon is undefined → fix the lookup or call site that passed bad data, not applyDiscount
• Follow existing code patterns and keep scope tight
• Keep temporary instrumentation that helps prove the repair until verification is complete, then remove it

3. Verify with matching evidence

A repair is only done when the evidence matches the report. Prove three things: (1) the original failure is gone, (2) the repaired path was actually exercised, and (3) nearby behavior did not regress.

Hand off to a tester — an isolated teammate that verifies the repair independently. See .claude/skills/test/teammate.md for how to spawn one.

Add or update a durable test in /fix when covering the bug clearly belongs in the codebase. Otherwise the tester owns verification.

4. Clean up

Remove temporary debugging artifacts once verification passes: throwaway scripts, temp logs, or ad hoc instrumentation. Keep durable tests and intentional logging.

GATE: Do not call the bug fixed until the evidence directly addresses the reported failure.

Issue

$ARGUMENTS

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: avibebuilder
• Source: avibebuilder/claude-prime
• License: MIT
• Homepage: https://claudeprime.avibebuilder.com

Install and usage instructions live in the source repository linked above.