AI Field Manual demo catalogue
Demo

Demo catalogue. These entries were written to prove the format; they are not an exhaustive or endorsed list, and their review dates are sample records. What demo means here

FM-09 Reviewed · v1.1Review due

Review a pull request with an agent that cannot change anything

A read-only agent proposes findings with file, line and failure scenario; a person verifies each one before it is posted.

You end with

A list of verified review findings, each with a concrete failure scenario, and a running count of false positives.

Level
Practitioner
Effort
35 min first time
Works in

Is it the right skill?

Use it when

  • Reviews are a bottleneck and you want a second reader.
  • The change touches code you know less well.

Skip it when

  • The diff contains secrets or customer data.
  • You would post findings without checking them.

Before you start

  • An agent that can run with read-only permissions.
  • The pull request's intent in one or two sentences.

Steps

Do not move on until the checkpoint holds.

  1. Start read-only

    Configure the session so the agent can read files and the diff but cannot edit, commit or push.

    Checkpoint A test write attempt is refused.

  2. Give intent and risk areas

    Say what the change is meant to do and which parts worry you.

    Checkpoint The prompt names the intent and at least one risk area.

  3. Ask for scenarios, not opinions

    Each finding must give file, line, a concrete input or state, and the wrong result.

    Checkpoint Findings without a failure scenario are discarded.

  4. Verify before posting

    Reproduce or reason through each finding yourself. Post only those you can stand behind.

    Checkpoint Every posted finding has your own confirmation note.

  5. Count false positives

    Record how many findings were wrong, per review.

    Checkpoint You know whether the agent is saving time on this codebase.

What can go wrong

  • Security Diffs can contain credentials committed by mistake. Scan before sharing.
  • Effect on people The person posting the review is accountable for it, and for its tone.

Revision history

Demo records, newest first.
VersionDateChange
v1.1Added the false-positive count.
v1.0First published.