You cannot read every pull request a model opens

Who this is for

• Engineering managers and tech leads reviewing code they did not write.
• Staff engineers who became the default reviewer for anything AI-generated.
• Anyone merging model-written code where a quiet mistake gets expensive.

The problem

• Four open before standup, three of them mostly generated. You skim, and the one that mattered is the one you skimmed.
• Generated code invents calls that read perfectly and do not exist.
• It shifts a boundary by one inside a loop that looks correct.
• It swallows an error the old path used to raise and hands the caller a success.
• Your checklist was written for human mistakes. It walks past all three.

What it does about it

• Triages every file: four buckets, and a count of the added lines you just committed to reading.
• Starts with the deleted lines: most quiet damage in a rewrite is in what was removed.
• Hunts ten failure modes: the ones generated code has and human code does not.
• Sorts findings four ways: including a bucket for scope creep, so it stops merging as a nitpick.
• Signs off on a scope: what was read, what was skimmed, what nobody could check.

Two minutes to first output

1. Paste the pack into Claude, ChatGPT or Copilot.
2. Add the ticket, the description and the diff.
3. Read the triage first.

What it will not do

• Confirm a call exists without the imports or the repository. It lists what it could not check.
• Review what you did not paste. If the diff is elided it says so up front.
• Run your tests.