AI-Agent Regression Debugging Prompt System

What it does for you

Find the one change your coding agent made that caused the bug, fix only that, and add a test that would have caught it. Bugs from coding agents usually start with a wrong guess one or two steps before the error you see, so fixing the line that errors often misses the real cause. These 8 prompts walk you from the symptom back to that guess.

What's included

• prompts.md: all 8 prompts in full, each with when to use it, what to give it, and what it should hand back
• prompts.json: the same 8 prompts in a format you can load into a script, slash command or git hook
• templates/incident-note-template.md: a fill-in template for a short incident note
• templates/regression-test-checklist.md: a checklist to make sure your new test would have caught the bug
• examples/example-walkthrough.md: a worked example of a payment retry bug that charged a customer twice, taken through the prompts step by step
• README.md: which prompt to use when, plus a quick start

How to use it

1. Open the coding tool that made the change (Claude Code, Cursor, Windsurf or Copilot Chat), so the prompt can read your repo, the diff and git history.
2. Pick the prompt that matches where you are. For a fresh bug, start with Prompt 1 to find the commit that caused it.
3. Paste the prompt, replace the parts in double curly braces with your real error, command and commit, then work through the next prompts in order: find the cause, fix only that, add a test, and check it before you call it fixed.

Sample prompt

Prompt 1 of 8, shown in full.

Prompt 1: Regression Bisection Without a Rewrite

Use when: Something that worked now doesn't, and you have a commit range to search but not a specific commit.

I have a regression. Symptom: {{exact error message, wrong output, or failing test name}}.
It reproduces with: {{exact command}}.
It last definitely worked at or before: {{commit SHA, tag, or approximate date}}.

Do NOT fix anything yet. Your only job right now is to find the one commit responsible.

1. Use `git log --oneline` between the last-known-good point and HEAD to list candidate
   commits touching files relevant to the symptom area.
2. For each candidate, use `git show <sha>` to read the actual diff; do not guess from the
   commit message alone, commit messages on agent-authored commits are frequently a summary
   of intent, not a reliable description of what the diff actually does.
3. Narrow to the single commit where the symptom-producing behavior was introduced. If you
   cannot narrow past 2-3 commits with the evidence available, say so explicitly and list
   what additional evidence (a specific test run, a specific log line) would resolve it;
   do not pick one and present it as certain when it isn't.
4. Once you have the commit, quote the exact hunk (file, line range, before/after) that is
   responsible, and explain in one paragraph the mechanism by which that hunk produces the
   symptom. If the hunk alone doesn't fully explain the symptom, say what else is needed and
   keep looking rather than stopping at a partial explanation.

Report format: commit SHA, the quoted hunk, the mechanism explanation, and your confidence
(certain / probable / needs more evidence) with the reason for that confidence level.

Delivery

Instant download after Stripe checkout. File: nightforge-regression-debugging-prompt-pack.zip

All sales are final once your download link is issued; refunds are given only for a billing error (wrong amount or double charge) or a corrupted file we can't fix or replace. See Terms.