Debug with AI using evidence
https://patrickz.ai/learn/debug-with-evidence/

OUTCOME
Prepare a reproducible bug report and verify one focused fix.

YOU NEED
A small project, a coding assistant, version control and a reproducible error. Use a practice branch.

1. Freeze the symptom
Write the exact steps, expected result and actual result. Save the error text and the relevant screenshot. Note the device or environment.

2. Ask for ranked hypotheses
Request three likely causes and the cheapest observation that would distinguish them. Do not accept a rewrite before the cause is tested.

3. Apply one focused change
Make a checkpoint, inspect the proposed change and rerun the reproduction. Keep unrelated cleanup out of the patch so the result remains interpretable.

4. Test nearby behavior
Try the original failure plus one normal and one edge case. Save the evidence and a short explanation of the cause, not just “fixed.”

PROMPT
Investigate this bug. Reproduction: [steps]. Expected: [expected]. Actual: [actual]. Evidence: [logs/files]. Rank likely causes and identify the smallest check for each. Implement the smallest supported fix, rerun the reproduction and test nearby behavior. Explain what remains unverified.

EXAMPLE / EXPERIMENT
A character falling through a platform might be a collision layer, spawn position or timing issue. A debug view of collision shapes can distinguish these before a large code change.

CHECK YOUR RESULT
[ ] Someone else can reproduce the original symptom.
[ ] The fix addresses an observed cause.
[ ] Original and neighboring cases pass.

IF IT FAILS
If the agent keeps changing unrelated files, stop and ask for the hypothesis and supporting evidence. Restore a checkpoint when an experiment makes the situation harder to understand.

MY RESULT / NEXT CHANGE