recipes
0/4 steps
Tutorial V5 · 8–10 minutes

PR QA review from one link

One pull request URL in, a per-criterion verdict table with evidence paths out. You confirm nothing gets posted.

skill not merged yet maturity: experimental
V5

Video recording in progress

Status: shipping, not yet merged

The PR QA review skill is on a branch, not on the skills repo's main branch. This page describes the flow as built so you know what is coming; there is no install command here yet, because one that does not resolve is worse than none. V4's recipe-cook is merged and available today — the review flow is one of its modes.

The flow

  1. One URL, matched to your checkout

    Exactly one pull request from metamask-mobile or metamask-extension. A URL for the other client is rejected before any checkout or runtime work — the review runs against a real build, so it has to be the right one. Given no URL, it reads the current branch's pull request.

  2. Criteria are extracted, then frozen

    Supply criteria and they are preserved verbatim, never merged with or overwritten by the pull request body. Supply none and they are extracted from it. Either way they are frozen before any runtime work starts, so the thing being proven cannot drift to match what turned out to be provable.

    Missing criteria are a finding, not a gap to fill

    A pull request with nothing to extract produces a hygiene finding. The skill does not invent criteria to have something to test.

  3. Proof runs on a real device

    The branch is checked out, the app launched under harness control, and each criterion exercised against the running build. Evidence lands in a bundle you can open — the same structure V2 walks.

  4. A verdict table you post yourself

    Output is a QA report plus a proposed review reply, per criterion, with evidence paths.

    It never posts. The reply comes back for you to read, edit, and submit. A machine-generated review that appears on a colleague's pull request without a human deciding to send it is not review — it is noise with your name on it.

    The verdict that matters most is the honest one: a criterion the surface genuinely cannot demonstrate is reported as needing a human, not quietly passed. A gesture-driven interaction with no automatable surface is the standard example — it gets classified, not faked.

Next: release validation → How to read the evidence it produces Back to the series