<< All versions

Skill v1.0.0

currentAutomated scan100/100
theam/facility/reviewing-to-standard
──Details
PublishedNovember 5, 2026 at 06:46 PM
Content Hashsha256:a3916e37978133aa...
Git SHAece5b14e8212
──Files
Files (1 file, 2.2 KB)
SKILL.md2.2 KBactive
SKILL.md · 52 lines · 2.2 KB

name: reviewing-to-standard description: Review code changes against this repo's STANDARD.md. Use when reviewing a PR, a diff, or another agent's work — enforces the review order, the comment quality bar, and severity discipline. metadata: version: "1.0.0"


Reviewing to standard

A review exists to change the implementer's next action. Everything else is noise.

Review order (stop-the-line first)

  1. Correctness and fit — the change does what was asked, no more. Scope

creep is a finding, not a bonus.

  1. Security and privacy — auth regressions, data exposure, secret

handling, injection paths, new attack surface. Verify the negative case; never accept a prose claim that a boundary holds.

  1. Maintainability — same concept reused not duplicated, clear ownership

boundaries, domain names, side effects visible, easy for the next caller and the next test.

  1. Standard compliance — STANDARD.md and its module sections were

followed. A missed module requirement (seeds, analytics, evidence, exposure) is a product-quality gap, not optional polish.

  1. Verification evidence — the right checks ran for the risk taken, and

skipped checks are named. node guards/run.mjs should be green.

The comment contract

Every finding carries four things: file:line, the exact risk, the smallest practical fix, and the standard it missed. Severity is explicit — Blocker / High / Medium — and separated from optional suggestions. Never post:

  • Style-only nits, unless they hide a real bug or future maintenance cost.
  • Vague asks ("clean this up", "add tests") — name the missing case, command,

or assertion instead.

  • Requests for broad rewrites when a narrow change closes the risk.

Leverage

For domains with real risk, dispatch the matching reviewer subagent from .claude/agents/ and consolidate its findings rather than re-deriving them. If the same problem has now appeared twice in reviews, the review is the wrong tool: propose the guard (guards/) that makes the third occurrence impossible, and say so in the review.

Verdict

End with Ready / Ready with follow-ups / Not ready, the checks you ran, and nothing else. You never approve or merge — that signature is human.

All versions