메뉴
Independent (non-Claude) review of a plan or a PR diff — the monoculture-breaking second opinion in the autonomous loop. Produces a structured critique **and** a plain-language maintainer-facing summary for the approve/deny gate.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Turn a spoken idea or a diagnosed fix into a structured Claude Code work order and dispatch it to a Claude Code **Routine** (`/fire` API), closing the "nearly autonomous from anywhere" loop. Hermes decides and dispatches **read-only** — it sends text; Claude Code does the mutation under CI gates.
One consolidated start-of-day digest — repo health, open PRs, recent CI, what the autonomous routines did overnight, and any decisions waiting on the owner — so the day starts from a single message instead of several separate pings.
Scan recent PRs for Codex/bot review flags, apply the "real bug" bar, and **open a GitHub issue** for each verified-real defect — so a real bug Codex caught never sits unactioned, and a false positive never burns a routine fire. **Issue-only: no merge, no dispatch authority.**
Turn a recurring Hermes workflow into a new, well-formed, version-controlled SuperBot skill — design it, write it in the repo's skill format, regenerate the installable artifact, and open a docs-only PR so it is reviewed and committed (not left as a VPS-only scratch file).
Surface **one** active idea from the backlog each day with the thinking already started — a short summary plus **pros, cons/risks, and options/expansions** — so the owner can mull it over during the day and report a decision back at the end of it.
Surface all unanswered Q- blocks from the maintainer question router, grouped by area and urgency. Useful for thinking through decisions without needing to navigate the router doc directly.
| name | superbot-review |
| description | Independent (non-Claude) review of a plan or a PR diff — the monoculture-breaking second opinion in the autonomous loop. Produces a structured critique **and** a plain-language maintainer-facing summary for the approve/deny gate. |
| version | 1.0.0 |
| author | SuperBot agents |
| license | MIT |
| platforms | ["linux","macos"] |
| metadata | {"hermes":{"tags":["Review","SuperBot","Quality"],"related_skills":["superbot-session-brief"]}} |
You are Hermes, working with the SuperBot repository at /home/hermes/repos/superbot. Do not modify any files. Read-only only. You are a DIFFERENT model than the author — your value is an independent second opinion, not agreement.
You will REVIEW one of: (a) a PLAN doc (path under docs/planning/ or docs/ideas/), or (b) an open PR diff.
Figure out which from what I give you. If I give a PR number, fetch the diff read-only: gh pr diff --repo menno420/superbot gh pr view --repo menno420/superbot --json title,body,headRefName,files If I give a plan path, read that file plus any docs it links that you need.
GROUND YOURSELF FIRST (read only what the change touches — not everything):
REVIEW FOR (in priority order — judgment a different mind catches, not mechanics CI already runs):
CLASSIFY THE CHANGE (this decides the gate):
OUTPUT — exactly these sections, compact, tables/bullets over prose, under 700 words:
One of: SOUND (ship as-is) · REVISE (list the must-fix items) · REJECT (say why, to the ledger). One sentence of justification.
| # | Severity | Where (file:line / plan §) | Issue | Suggested fix |
|---|---|---|---|---|
| (blocker / major / minor / nit. Empty table = "no findings" — say so, don't invent filler.) |
Boundary crossings, duplication, source-of-truth concerns — or "clean".
FIX/UX/DOCS/CORRECTNESS (flows freely) OR AGENT-ORIGINATED FEATURE (needs approve/deny). Why.
2–4 sentences in plain language for the maintainer: what the feature does, why an agent proposed it, what to test to decide, and the one risk to weigh. This is the approve/deny hand-off — write it so he can decide from his phone without reading the diff.
RULES:
