Hyper PR Review

A finding is a hypothesis, not a deliverable. As of Aug 2026, frontier reviewers catch only 15-31% of what human reviewers flag, so recall is a lost cause and precision is the entire game. Precision comes from verification machinery, not from better prompting (prompting-only noise control has published evidence of outright failure). This skill's edge over hosted review bots is that it runs where the code executes: every candidate finding faces its cheapest disproof before it is reported, and execution adjudicates whenever the code can run.

Position in the toolbox: this skill is you conducting the review. cross-model-review is dispatch and consumption mechanics for a different model's review; run it as an additional lane, not instead of this. super-good-pr owns the PR body standard this skill audits against.

Establish the Scope

For a GitHub PR, live state is authoritative:

gh pr view <n> --json number,title,body,baseRefName,headRefName,headRefOid,isDraft,mergeable,reviewDecision,url
gh pr diff <n> --name-only && gh pr diff <n>
gh pr checks <n>

Bracket the capture: re-read headRefOid after gh pr diff returns. If it moved, the diff you hold is already stale; re-capture before reviewing.

For a local branch, fetch first, then three-dot, and enumerate untracked files (git diff silently omits them, so a brand-new file can escape review entirely):