Review an AI Agent's Work Like an Auditor

Prompt

Here is the log of actions taken on a task, and the final summary that was produced.

Task: {{task}}
Log: {{log}}
Summary: {{summary}}

Review it as an auditor would:
1. List every action that changed something, and say whether it can be undone.
2. Point out anything in the summary that the log does not actually support.
3. Identify the facts the work depends on that were never verified.
4. Name anything done that was outside the original task.
5. Say what you would check by hand before relying on this.

Do not restate the summary. Look for the gap between what was claimed and what the log shows.

Fill in these placeholders

• {{task}}
• {{log}}
• {{summary}}

How to use it

Paste the actual action log, not just the closing summary. The whole point of this prompt is comparing the two, and with only the summary there is nothing to compare against.

Example use case

Reviewing a run that reorganized a set of records, to confirm which changes were real and which were only reported.

Why this prompt works

An agent's summary of its own work is another prediction: it describes what a successful run would plausibly look like, which is not the same as what happened. The log is evidence. Asking specifically for claims the log does not support turns a vague sense of unease into a list of things to check.

Common mistakes

• Reviewing the summary alone, which is exactly the document most likely to be wrong.
• Accepting this review as final. It is a second opinion from the same kind of system, not verification.
• Skipping the manual checks it recommends, which is usually where a real problem would have surfaced.

Variations worth trying

• Ask which single action would be hardest to undo, and check that one first.
• Ask what a skeptical reviewer would suspect had been skipped.
• Ask it to reconstruct the outcome purely from the log, then compare that with the original summary.