What to check before approving an AI implementation plan

A checklist for reviewing an AI implementation plan before you approve it, applied to two fictional plan versions with different scope and risks.

Use a plan with a known gap

The Agent Review Pack contains plan-v1.md, plan-v2.md and review-checklist.md. These are fictional documents written for this exercise. The first plan leaves questions about scope and verification; the revision makes the intended work more explicit. Download both so the checklist has a concrete object.

Four passes through the plan

Goal. Can you restate the user’s desired result from the plan alone? If it adds work that was not requested, name that expansion. Scope. Does it identify files and external systems that will change? “Update the app” is not a boundary. Dependencies. Are unavailable APIs, permissions or assets acknowledged? Verification. Does the plan say what observation would show the result works, including a failure case?

Use a small review note: Question — evidence in plan — decision needed. In the sample, a plan can mention an export but omit what happens to local images; the reviewer should ask for an export check rather than assume success. The checklist is a way to make a decision visible, not an automatic score.

Approve the decision, then inspect the result

Approval of a plan does not approve unknown implementation details. Ask for a report linking changes and checks after the work. If a plan changes, compare revisions before accepting the next one. Aero Markdown is a reader for these local documents; it does not judge or approve agent plans. Its Mac app remains unavailable to download.

For a Claude Code-centered reading flow, see the plan guide. No measured reduction in errors is claimed.