review-round
Skill5 starsChanged 31 days agoMITView on GitHub
Run the codified reviewer loop on a diff or PR: author adversarial per-agent prompts pointing at the riskiest spots, launch whichever of code-reviewer, deploy-guard and pr-validator the user picks when asked, in parallel, bring the findings to the user and let them choose what is worth fixing, reproduce by execution each one they pick before fixing it, prove each fix with a non-vacuous regression test, and post the PR verdict comment that /release gates on. Use before shipping, or when the user says 'review round', 'run the reviewers', 'full review pass', or wants the pre-release review loop.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- glslang/windbg-mcptheir own version, first seen Nov 6, 2026
- MortisDei/hermaeustheir own version, first seen Oct 8, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- glslang/windbg-mcp★ 21same name
- MortisDei/hermaeus★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository2 other files it gives its agents.
- pr-checkpointSkill
- releaseSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 74
Read the Initiative field of each.
- Very specific: line 84
When the user hands you an explicit path, use that path.
- Distinctive: line 92
Write each reviewer a prompt aimed at the riskiest spots in this diff, naming what it must try to break.
- Distinctive: line 98
Start every chosen reviewer in one message so they run in parallel, and say which ones were skipped and why.
- Distinctive: line 116
Check the FINAL value once, after every step that can replace it (a correction, an override), never each source: a source check refuses a …
- Distinctive: line 157
Every finding the user picks to fix becomes a work item on the project's todo list, in the order they have to be taken, and so does every …
- Very specific: line 165
Never merge, deploy, or push to the base branch from this skill.
- Distinctive: line 167
Never post CLEAN while a Critical/Blocker is not Fixed, Declined or Deferred.
- Very specific: line 168
Never post CLEAN while the release security scan has an unfixed, unwaived [NEW] finding at the reviewed head (step 8).
- Distinctive: line 170
Never bind a verdict to a SHA you did not actually review and test.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Review round
The codified review->fix->re-verify->verdict loop. Reviewer agents driven by hand-authored adversarial prompts catch Criticals that authoring and testing miss; this skill makes the discipline repeatable instead of re-invented per PR. It drives the companion review agents (code-reviewer, deploy-guard, pr-validator), published separately as the claude-review-agents pack.
Model note: the load-bearing work here is judgment (adversarial prompt authoring,
adjudicating findings). The pin is model: inherit, so this runs on the session
model; choose a strong reasoning model for it with /model if the session is on a
lighter one.
Two modes
- Pre-release mode (default): the full loop, steps 1-8. Step 3a puts the findings in front of the user and waits. What they choose gets fixed or formally declined before the verdict posts; the rest is recorded. "The full loop" has never meant "fix everything the reviewers name", and reading it that way cost a working day once.
- Checkpoint mode (invoked from
/pr-checkpointstep 6): steps 1-3 and 8 always; the fix loop (4-7) is optional and NOTHING blocks the PR. The checkpoint stays advisory by design; the hard gate is/release. Its step-8 verdict saysCHECKPOINT-ADVISORY, neverCLEAN, so the mode is visible to the gate that reads it later. See step 8 for why the two must not share a word.
The loop
Eight steps. Step 8 posts the verdict comment that /release gates on. It is easy to
drop because it comes after the work feels finished. It is not optional, and it cannot
be inferred from a transcript later.
1. Scope
Establish what is under review: git diff "$BASE"...HEAD for a branch,
gh pr view / gh pr diff for a PR. Record the reviewed HEAD SHA
(git rev-parse HEAD on the branch); the verdict in step 8 binds to it.
Stacked branch (base not main)? First run this folder's stacked_base_check.py
--repo <worktree>; 1: merge base, re-test; 2: stop.
Did a helper build it? Check it against the user's pick FIRST. Before any reviewer runs, read the builder's report for every decision it made itself ("a call I made", "I decided", "instead", "chose", "not X but Y") AND each design line of the brief you gave it (a builder obeying a rewritten brief reports no choice), and compare both, word by word, with the pick or the recorded decision. Anything that narrows, widens or replaces it goes to the user now, in one question: a reviewer must never be the first to notice.
1a. Choose the reviewers before launching any of them
Decide which review agents this repo needs, and settle it before step 2.
Read ~/.claude/skills/review-round/reference/1a-choose-reviewers.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
Find the spec that belongs to your initiative
A repo hosts more than one initiative over its life. The first writes REQUIREMENTS.md
and SPEC.md at the repo root, and a later one writes its own under
docs/<initiative>/. So the copy at the root is not always the one you want, and
grounding on the wrong initiative's spec produces work that is internally consistent
and aimed at the wrong system, which review does not catch.
Say which initiative this work is for, then resolve the file instead of assuming the root copy:
- List the candidates: the repo root copy, every
docs/*/copy, and any other location this skill already tells you to check. - Read the
Initiativefield of each. That field decides, never the directory name, which is a slug somebody chose and can be stale or wrong. - Exactly one names your initiative: use it, and say which path you used.
- None names it: STOP and tell the user what you found and where. Do not fall back to the root copy.
- More than one names it: STOP and name every path. Two files claiming one initiative is a thing to report, not to settle by picking one.
- A candidate carries no
Initiativefield: say so. Where it is the only candidate you may use it, and you must say you read a spec with no recorded owner.
When the user hands you an explicit path, use that path. Check its Initiative field
the same way, and on a mismatch stop and say whether a file for your initiative sits
somewhere else.
This round reviews a diff, not a specification, so it opens these files only when a finding is anchored to one. A finding anchored to the wrong initiative's spec is worse than an unanchored one, because the citation makes it look checked.
2. Author the adversarial prompts
Write each reviewer a prompt aimed at the riskiest spots in this diff, naming what it must try to break.
Read ~/.claude/skills/review-round/reference/2-adversarial-prompts.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
3. Launch the chosen reviewers in ONE message
Start every chosen reviewer in one message so they run in parallel, and say which ones were skipped and why.
Read ~/.claude/skills/review-round/reference/3-launch-reviewers.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
3a. Every mode: take the findings to the user BEFORE fixing any of them
Report every finding, then stop and let the user choose what gets fixed.
Read ~/.claude/skills/review-round/reference/3a-findings-to-the-user.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
4. Reproduce before fixing
Reproduce each chosen finding by running the failing input through the real path before any fix is written.
Read ~/.claude/skills/review-round/reference/4-reproduce.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
Before writing any fix, list by command every path to the finding's RESULT (step 5a) and put the list in the fix commit message: a finding names one site, its result usually has others. Check the FINAL value once, after every step that can replace it (a correction, an override), never each source: a source check refuses a value a later step would have fixed.
5. Verify suggestions against the authoritative gate
A reviewer's proposed fix is a hypothesis. Check it against the authoritative release gate or the code before applying (the red-link lesson: one linter's complaint was another gate's by-design zero). Declining a finding with a stated reason is a legitimate outcome and goes in the verdict comment.
5a. Prefer the fix that removes, and flag the fix that adds
Close a finding by deleting or narrowing where you can; name every added piece for the re-review; and when a finding names a MECHANISM, list every path the changed function has to that result.
Read ~/.claude/skills/review-round/reference/5a-prefer-removal.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
6. Every fix gets a non-vacuous regression test
Each fix gets a test that fails without the fix, proven by breaking the guard or by contrast, and the suite total must rise by exactly the tests added.
Read ~/.claude/skills/review-round/reference/6-regression-tests.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
7. Re-run
Touched suites first, then the FULL suite. A subset of suites is not "green".
8. Post the verdict comment
Post one PR comment with the per-reviewer counts, what was fixed or declined, and the last-line verdict that /release reads.
Read ~/.claude/skills/review-round/reference/8-verdict-comment.md now, in full, and follow it. It is this step's full text; the line above is only a summary and is not enough to act on.
Background-session note
The verdict-then-merge flow runs gh pr comment / gh pr merge, which the
auto-mode classifier can block on agent-authored PRs. A transient stage-2
classifier error gets ONE plain retry; a sustained block means asking the user to
switch the session to manual mode, not fighting it.
The work items go to the todo list
Every finding the user picks to fix becomes a work item on the project's todo list, in the order they have to be taken, and so does every one they defer rather than reject. A deferred finding with nothing on the list is indistinguishable from one nobody found. The agents themselves write nothing: a subagent cannot ask the user anything, so the write belongs here, in whoever spawned them.
Guardrails
- Never merge, deploy, or push to the base branch from this skill; the ship verb is
/release. - Never post CLEAN while a Critical/Blocker is not Fixed, Declined or Deferred.
- Never post CLEAN while the release security scan has an unfixed, unwaived
[NEW]finding at the reviewed head (step 8). - Never bind a verdict to a SHA you did not actually review and test.
- In checkpoint mode, nothing here refuses, blocks, or rolls back the PR.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

