Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Other skills

cross-model-review

Use when a reviewed change is about to land on a shared branch, be released or be deployed, and when the user asks for a second opinion on a change from a different model or a different AI vendor.

NoobyGains · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add NoobyGains/godmode --skill cross-model-review

About this skill


name: cross-model-review description: "Use when a reviewed change is about to land on a shared branch, be released or be deployed, and when the user asks for a second opinion on a change from a different model or a different AI vendor."

Cross-Model Review

Models from one family share blind spots: the same training habits, the same favourite mistakes, the same things they never question. Before a change lands, one reviewer from a different model family than the author reads it once, narrowly, from a brief file. One focused outside read catches what a whole team of same-family reviewers will agree to miss.

The Prime Directive

NOTHING LANDS WITHOUT ONE READ-ONLY REVIEW FROM A DIFFERENT MODEL FAMILY, OR A RECORDED REASON WHY NOT

No exceptions. No workarounds. No shortcuts.

When to Use

Required:

  • Before merging a feature branch into the main branch.
  • Before a release, a tag or a deployment.
  • Before landing changes to security-sensitive code, data migrations, public APIs, build or release tooling.
  • When the user asks for a second opinion, an outside review or a review by another vendor's model.

Skip:

  • Work that has not passed godmode:code-review yet. This review comes after, not instead.
  • Throwaway spikes and local experiments that will not land.
  • Changes where the user has said explicitly that they do not want an outside review; record that they said so.

The Protocol

  1. Confirm the entry state. godmode:code-review has passed with no open Critical or Important findings, the suite is green, and the work is committed. Note the base and head commits.

  2. Identify the author's family. The family is the vendor's model line (for example Claude, GPT, Gemini), not the size or tier. A different tier of the same family does not count as a second family.

  3. Pick the reviewer. Choose a reachable model from a different family. How to reach a second model family on your harness (which command, which window, which plugin, which flags keep it read-only) is in the harness reference of godmode:using-godmode; see the harness reference for how to reach a second model family. Pick the cheapest model of that family that is good at code review at a medium effort setting; do not pick a premium tier or a fast mode that burns the user's allowance for one review (see godmode:model-routing).

  4. Write the brief file. Copy the review brief template to a scratch or git-ignored path and fill it in:

    • the exact commit range and the exact files to read, each with a one-line reason;
    • three to seven exact questions, each one checkable from the code ("Can parseConfig return a partially populated object when the file is truncated?", not "Is the config code good?");
    • the requirements the change must meet, or the path to them;
    • the constraints: read-only, no sub-agents, no web browsing unless a question needs it, one round, the report format, and where to write the report.

    Narrow beats broad. A reviewer given the whole repository and "review everything" returns a long list of style opinions; a reviewer given six files and five questions returns the bug.

  5. Send it once. Hand the reviewer the path to the brief, not a pasted copy of your session. Ask for the report as a file at the path the brief names. Do not chat with the reviewer during the round and do not send follow-up rounds; if the report raises a question only the reviewer can answer, that is a new review with a new brief, and it needs the user's agreement.

  6. Wait for the report and read all of it. If the reviewer fails (it times out, it is out of quota, it edits files), record the failure, discard any edits it made after looking at them, and go to step 9.

  7. Reconcile with evidence. For every finding, before changing anything:

    • reproduce it: run the input, read the code path, write the failing test;
    • classify it as confirmed, rejected (with the evidence that refutes it) or needs the user's decision;
    • fix confirmed findings test-first, with godmode:test-first, and re-run the verification yourself.

    The outside reviewer's verdict is an input, not proof, in both directions. "Looks good" from another family does not replace your own verification, and an alarming finding is not real until you have reproduced it. Use godmode:review-response for the evaluation.

  8. Check the fixes yourself. Do not send a second round to the outside reviewer for the fixes. Run targeted tests for each fix and the full suite, and record the output.

  9. Fallback when no other family is reachable. If no other family is installed, reachable or within quota, dispatch a fresh reviewer from the same family with no session history, using the same brief file, and record "fallback: same family, because ..." in the outcome. Say this plainly to the user; a same-family review is weaker, not equivalent.

  10. Record the outcome. In the pull request description, the progress ledger or the release notes:

    Cross-model review: <family and model> via <how it was reached>, <date>
    Range: <base>..<head>   Brief: <path>   Report: <path>
    Findings: <n> confirmed and fixed, <n> rejected with evidence, <n> for the user
    

    If the review was skipped, record who decided and why. Then hand on to godmode:merge-protocol.

Writing Good Questions

Weak questionStrong question
"Is this secure?""Can a user without the admin role reach DELETE /projects/:id through the batch endpoint in routes/batch.ts?"
"Any bugs?""What does retryWithBackoff do when the callback throws synchronously on the first call?"
"Is the migration safe?""Does migration 0042 lock the orders table for longer than a single statement on PostgreSQL, and is it reversible?"
"Check the tests.""Which branch of applyDiscount has no test that would fail if it were deleted?"

Aim questions at the places a same-

Read the complete source ↗