human-review-guide

human-review-guide — a local guide for the human review gate

Core principle: the human gate covers only what the gates cannot see

Code and rules are already reviewed by agents and gated: crew self-review, the owner's agent review, CI, lint, guard tests, knowledge-base rule checks, a repo's PR review tool. The human gate exists only for what those gates cannot see or did not cover. Two kinds of thing qualify:

1. Not covered — no gate looked at it in this diff.
2. Only a human can judge it — intent and scope, public contract names, architecture trade-offs, whether the tests test the right thing, a live comparison with a design or prototype.

Everything else is dropped from the guide, and the human is never asked to redo it. The guide also records the gate gaps: things that should have been machine-checked but no gate covers, each with a suggested new gate or follow-up ticket, so a human catches it once and a machine catches it next time.

This skill produces the guide that makes that review fast.

Structure follows addyosmani/agent-skills code-review-and-quality (five axes, tests first, severity labels, common rationalizations).

Who writes it

The agent in charge of the PR(s) writes the guide itself, whichever session briefed, reviewed and drove them (normally the captain). It holds the context: why each choice was made, the review findings, the decisions taken since. Do not hand the guide, or its HTML, to a new crew, a side session or a fresh subagent; it lacks that context and produces a shallow guide. If you have rolled over or compacted, reload context first (your handoff, gh pr view N --comments, the commit history, the review findings) and then write it yourself.

LOCAL ONLY — never post the guide

The guide is for the operator. Never post it, or any part of it, on a PR, issue, tracker ticket, chat channel or any other shared place. Not as a comment, a description edit, a review body or a status update.