The AUDIT Prompt

I had Codex audit a plugin I built, and it came back with a huge list of fixes. AI is good at building. A second AI is better at catching what the first one missed. The hard part is turning a long audit into something you can actually finish.

This is how I do it, in three steps:

1. Run the audit
2. Rank the list
3. Fix one batch at a time

Before you start

Use a different AI to audit than the one that built your work. Built it in Claude? Audit it in ChatGPT or Codex, or the other way around. At minimum, use a temporary or incognito chat so the reviewer doesn’t remember the work.

Remove passwords, API keys and customer data before you paste anything. If the audit finds one already in your work, change it at the source first (a new password or a new key), then remove it. Deleting it from the file doesn’t make it safe once it’s been exposed.

Give the AI something it can read. Website: the URL plus screenshots. Automation (Zapier, Make, n8n): screenshots or a list of the steps. Code: if you use Codex or Claude Code, point it at the whole project instead of pasting files, so it can catch places where one file contradicts another.

Step 1: Run the audit

In the reviewing AI, paste this prompt, fill in the brackets, then attach or paste your work.

You are a skeptical senior reviewer. Your job is to find real problems, not to praise.

WHAT I'M SHARING: [plugin / code / doc / automation / website / process]
WHAT IT'S SUPPOSED TO DO: [1–2 sentences]
WHO USES IT: [me / my team / customers]

Review it for:
1. Things that are broken or wrong
2. Things that are missing, compared to what it's supposed to do
3. Risks: security, privacy, data loss, cost, or anything that could embarrass me
4. Inconsistencies: places that contradict each other
5. Things that are confusing or hard to use

Rules:
- Point to the exact spot for every issue (file, line, section, or quote).
- Only report what you can see. If you're guessing, label it "Unverified."
- Don't invent issues to fill the list. A short list of real problems beats a long list of nitpicks.
- Don't rewrite anything yet. Just diagnose.

Output a table with: # | Issue | Where | Why it matters | Severity | Suggested fix (one line, no rewrites)

Severity: Critical = broken or unsafe now. High = will cause problems soon. Medium = worth fixing. Low = polish.

Show every Critical and High issue. If there are more than 15 Medium and Low issues, show the top 15 and tell me how many more you found.

Then finish with:
- The 3 issues I should fix first, and why
- What's working well that I should NOT change.

Step 2: Rank your list

Go back to the AI that built your work. Paste the full audit from Step 1, then paste this underneath it:

Here's an audit of your work from another reviewer. For each item, tell me whether you agree, disagree, or need more info, and explain briefly. Judge each item on its merits. Don't defend your earlier work. For every item you disagree with, quote the evidence. I'll decide those, not you.

Then group the items you agree with into:
- MUST FIX (broken or risky)
- SHOULD FIX (makes it meaningfully better)
- NICE TO HAVE (skip for now)

Break MUST FIX into small batches, grouping related fixes together. Don't change anything yet. Show me the ranked list, the batches, and the items you disagreed with.

Read the ranking before going on. Look hardest at the items the builder disagreed with. Those are often the most important ones. If you’re not sure who’s right, paste the disagreements back to the reviewing AI and let it answer. If you disagree with where something landed, say so now. Moving an item is easy at this point and hard once the changes are made.

Step 3: Fix one batch at a time

Before each batch, save a copy of your work (or commit it, if you use Git) so you can undo a batch that breaks something. When the ranking looks right, paste:

Fix batch 1 only. Change only what this batch covers. No other improvements.

Don't fix an issue by deleting, skipping, or loosening a test, check, or validation. If that's the only way, stop and tell me.

When you're done, show me exactly what changed. Then check that it works and show me the actual result (output, screenshot, or test run), not a summary of it. Wait for my OK before starting the next batch. When MUST FIX is done, ask me whether to move on to SHOULD FIX.

Check each batch yourself before you reply “OK, next batch.” If something broke, say what you saw and have it fixed before moving on.

Tip: when the fixes are done, run Step 1 again in a fresh chat, without mentioning the first audit. A second audit shows whether the fixes worked or caused new problems.