Automation & Agents

Getting AI to Audit an AI-Written Plan (With a Copy-Paste Review Prompt)

AI-written plans always look complete — self-consistent logic, clean formatting — until you're two weeks into execution and find a rule that contradicts itself. The fix isn't rereading it yourself. It's handing it to an AI that never saw how it was written.

09/26/2026 ｜ 2 min read

One Core Idea

The person who wrote it can’t find their own blind spots. Neither can an AI. Ask the same AI, in the same conversation, to “double-check the plan you just wrote,” and it inherits every assumption and reasoning habit from that thread. What it finds is typo-level. A useful review needs two conditions:

1. A clean (fresh) context: the reviewer sees only the finished document, never the reasoning behind it.
2. Forced failure scenarios: reject vague “you could consider adding…” suggestions — every finding must trace a concrete “under what input/state does this break.”

How To Do It

Step 1 | Save the plan as a standalone document

Paste the AI’s full plan into a new file or document — don’t leave it sitting in the chat. The reviewer should only ever see this finished artifact.

Step 2 | Open a genuinely new AI session

The key word is “new”: a fresh chat window, or a separate subagent if your tool supports one. Don’t hit “continue” in the original thread. If your tool allows it, use a stronger reasoning mode for the review pass.

Step 3 | Paste this review prompt

You are an independent second-opinion reviewer, unrelated to whoever wrote this plan.
Review it as an unsparing outside observer. Find problems that will actually bite during
execution, not copy-editing suggestions.

[One-line context]: (e.g. solo execution, X hours/day, goal is Y)
[Full plan]: (paste here)

Review across five dimensions:
1. Process soundness: any stalls, ordering issues, or hidden dependencies ("A only works once B exists")?
2. Assumption validity: which assumption is most likely to be false?
3. Gaps: given the end goal, what's missing that would materially affect the outcome?
4. Over-engineering: what doesn't deserve effort at this stage?
5. Single points of failure: where is the system most fragile?

Acceptance criteria: every finding must include a concrete failure scenario (what input/state
→ what breaks). Do not accept vague "consider adding..." items. If a section is genuinely
sound, say so plainly — don't manufacture criticism.

Report format: (1) one-line verdict (ready to execute / needs revision) (2) findings ranked by
severity, each as problem → failure scenario → one-line fix (3) up to 3 "don't do this yet"
cuts.

Step 4 | Write the findings back into the plan

Go through each finding one at a time: accepted fixes get written back into the plan document; rejected ones get a documented reason. Tag every edit with a date and source (e.g. “2026-XX-XX, second-opinion review fix”) so future-you knows which clauses were already patched.

Step 5 (advanced) | Treat the “cut this” list as the real prize

The “don’t do this yet” items are often worth more than the “add this” items — they let you drop work that looks diligent but returns nothing right now.

Pitfalls You Will Hit

1. Asking the same AI, same conversation, to self-check — and believing it counted as a review. Same-context “self-checks” only catch surface issues. Structural dead-locks (a rule that contradicts its own execution order, a hidden dependency) survive almost every time, because the model thought it was reasonable when it wrote it, and still thinks so on a reread. Step 2’s clean context is the only fix — there’s no shortcut.
2. A review prompt with no acceptance criteria comes back with ten polite, toothless suggestions. Without the “concrete failure scenario” rule, an AI reviewer defaults to listing safe, low-friction ideas. That one rule filters out the noise; everything that survives is worth acting on.
3. Accepting every finding wholesale (or dismissing all of them). The reviewer can also be wrong — it never saw your full context. The right move is a finding-by-finding call: accept and write back, or reject and record why. You’re the one who signs off; the AI is staff, not the decision-maker.

When You Get Stuck, Ask an AI

Paste this into any AI:

I'm a beginner following a guide on "using AI to review an AI-written plan."
I'm stuck at: (describe the step)
Tool I'm using: (ChatGPT / Claude / other)
What just happened / full error message: (paste verbatim)
Explain in plain language, step by step, exactly what to click or paste next.

One-Line Summary

A plan looking polished proves nothing. What proves something is an AI that never saw how you thought about it, working out exactly how it would fail.