scrutinize
Skill14 starsChanged 49 days agoMITView on GitHub
Multi-agent review and remediation; correctness, security and simplification lenses, then fixes criticals and highs. Use to review a diff or PR before shipping.
7 other projects use a version of this.
Adopted by
Most recently taken up by
- rcdelfin/agentkittheir own version, first seen Aug 28, 2026
- roborracle/agent-workflow-kittheir own version, first seen Aug 11, 2026
- autsachi/codex-skillstheir own version, first seen Aug 11, 2026
- neronain/ClaudeSkills-Neronaintheir own version, first seen Aug 9, 2026
- JirawatHQ/agis-methodtheir own version, first seen Jul 5, 2026
- thananon/9arm-skillstheir own version, first seen Jun 14, 2026
Used with
Projects with it also use management-talk and post-mortem.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
thananonSkillmanagement-talk
Rewrite engineer-to-engineer content for engineering-org leadership (VPs, directors, PMs, release managers, execs in an engineering-savvy company) and shape it for the channel it is going to — JIRA comment, Slack post, async standup line, email, or meeting talking-points. Trigger when the user asks to write/rewrite for management / exec / VP / director / PM / release manager, asks for an "executive summary / leadership update / status update", says "make this less technical / less jargony", or asks for a slack / email / standup / meeting version of work originally written engineer-to-engineer.
Also in 3 projects that use this
andreluizwebSkillpost-mortem
Guia interativo para criar um Post-Mortem de incidente, coletando informações uma pergunta por vez
Also in 3 projects that use this
Other versions7 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 7
- thananon/9arm-skills★ 3.2ksame name
- neronain/ClaudeSkills-Neronain★ 3same name
- rcdelfin/agentkit★ 2same name
- autsachi/codex-skills★ 0same name
- JirawatHQ/agis-method★ 0same name
- jpsweeney97/claude-code-skills★ 0same name
- roborracle/agent-workflow-kit★ 0same 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 repository24 other files it gives its agents.
- accelerateSkill
- almanacSkill
- autopsySkill
- blueprintSkill
- certifySkill
- chronicleSkill
- confluenceSkill
- conjureSkill
- deploySkill
- divineSkill
- inscribeSkill
- jiraSkill
- knowledge-graphSkill
- magicSkill
- manifestSkill
- orchestrateSkill
- portalSkill
- sealSkill
- sentinelSkill
- statuslineSkill
- transmuteSkill
- unravelSkill
- wardSkill
- weaveSkill
The file
We split this skill into the rules it gives an agent
3 rules. 0 are common, 1 is distinctive, and 2 are very specific to this project.
0
Common
Most of the 6 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 23
If REVIEW.md exists at the repository root, Read it first.
- Very specific: line 89
Round ledger: Before the first fix, write .workspace/shared/diffs/<ref>-rounds.md in the format of lore/deep/review.md#round-ledger.
- Distinctive: line 94
Run the affected test, then the full suite.
“Common” is measured: kept by at least 6 in 10 of the 6 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/scrutinize — Multi-Agent Review & Remediation
Bundled command: if
kgis not found, run it as${CLAUDE_PLUGIN_ROOT}/tools/kgand give subagents that full path. If that is missing too (Claude chat ships no plugin tools), skip thekgsteps and search the code directly.
Review a code change with three specialist lenses (inline when small, parallel agents above), consolidate the findings, then remediate in at most three rounds. Review and fix are one loop.
Completion Signal
"Scrutinize complete. <N Important, N Nit, N Pre-existing; N fixed, N closed with reason>. Run /certify to verify clean state."
Dialect
The review follows lore/review.md. Every finding carries a tag (Important, Nit, Pre-existing) and file:line. Nits are capped. Remediation runs at most three rounds, each after the first on the diff since the last round only, with a round ledger at .workspace/shared/diffs/<ref>-rounds.md. The count line is the first line of the report and of the summary. The findings follow it.
If REVIEW.md exists at the repository root, Read it first. Its values (what failure means, skip paths, nit cap, round cap, risk tiers) override the defaults, and its text goes into every lens prompt. Remove its skip paths from the diff before the lenses run.
Effort
Scale effort to the change size. A large or security-sensitive change warrants /effort high (xhigh for a sprawling diff, max only as an opt-in). See lore/models.md.
Autonomy — approve the plan, then run
Phase 1 runs autonomously. Do not ask before reads, searches, kg query/blast, or read-only git diff/status. This skill pre-approves kg query/blast and the writes under .workspace/shared/diffs/. The Phase 2 fix edits go through the normal permission prompt unless the user runs in auto mode.
The SCRUTINY REPORT (Phase 1, step 7) is the single approval gate: end your turn there and wait. Once approved, Phase 2 runs the fix rounds without gating on intermediate reads. Re-gate only on real side effects: the fix Edits and a decision to close an Important finding without a fix. See lore/autonomy.md.
Phase 1 — Review
- Collect the scope and write the diff once.
<base>is the first word of$ARGUMENTSwhen given, otherwisemain. Run each as its own plain command so it matches the pre-approval:
Ifgit diff <base>...HEAD --shortstat git diff <base>...HEAD --name-only git diff <base>...HEAD > .workspace/shared/diffs/review.patch.workspace/exists but.workspace/shared/diffs/does not, create that directory first (themkdirprompts). With no.workspace/at all, do not create one. Rungit rev-parse --git-dirand write the patch to<git-dir>/magician-review.patchinstead, outside the tracked tree (that write also prompts). Note the patch path. Drop theREVIEW.mdskip paths from the file list. A small diff (3 files or fewer, 300 changed lines or fewer, by--shortstat) needs no patch file. Small diff: run the lenses inline. A High-tier change dispatches its lenses fresh at any size. Below the threshold, run each lens in the main loop with
${CLAUDE_PLUGIN_ROOT}/agents/<lens>.mdas its checklist, one lens at a time, and make no lensAgentcall (lore/fold-first.md). Its re-review rounds run inline too. The gatekeeper stays fresh.Larger diff: dispatch the lenses in one message. Fork or fresh follows
lore/deep/review.md#fork-or-fresh. When the diff is already in context (about 40k tokens or less), each lens is a fork (subagent_type: "fork"). The fork prompt names the lens and tells it to Read${CLAUDE_PLUGIN_ROOT}/agents/<lens>.md. Otherwise use these fresh subagent types, which the plugin registers:magician:reviewer: correctness and edge casesmagician:sentinel: security (OWASP, secrets, injection into app code)magician:simplifier: over-engineeringmagician:guardian: add this lens when the diff touches the agentic surface (agents, hooks, tools, skills, prompts, untrusted input)
With
--lens <name>in$ARGUMENTS, dispatch only that lens (the projectsimplifyskill calls/scrutinize --lens simplifier).Context contract: each
Taskprompt is self-contained, because the agents see none of this conversation. Pass the patch path from step 1 and the changed-file list, never the pasted diff. Add the goal ("review this change for<lens>"), the conventions and lore in play, theREVIEW.mdtext, the ledger path on a re-review, and the output format below (lore/subagent-context.md). OnNEEDS_CONTEXT, add the missing input and re-dispatch.Each agent returns findings as:
SEVERITY: Critical | High | Medium | Low TAG: Important | Nit | Pre-existing FILE: path:line ISSUE / VULNERABILITY: <what> FIX: <remediation>Finders optimize for recall. Consolidation optimizes for precision. Never tell a lens to be conservative, to skip nits, or to report only high severity. Models follow that literally and withhold bugs. Ask for full coverage with confidence and severity attached, then filter in steps 4 and 5.
- Collect all findings.
- Deduplicate. Merge the same issue from several lenses into one (note all sources). Drop findings the code refutes and findings without
file:line. - Prioritize. Critical, then High, Medium, Low. Keep at most the nit cap (default 5) and note how many nits you withheld.
- Present the consolidated report:
=== SCRUTINY REPORT === N Important, N Nit, N Pre-existing; 0 fixed, 0 closed with reason. Critical: N | High: N | Medium: N | Low: N [Critical] Important FILE:LINE — Issue (Source: reviewer/sentinel/simplifier) Fix: remediation steps ... Approval gate (AskUserQuestion). Present the report, then ask how to proceed through AskUserQuestion (never bare prose):
- Fix Important findings now (default): proceed to Phase 2.
- Discuss first: talk through the findings before any fix.
End your turn at the tool call. Wait for the choice before you remediate. Treat a free-form "yes / approved / looks good" as Fix Important findings now. When the request itself already says to fix (for example "review it, then fix the Important findings"), that is the answer: show the report and go on to Phase 2 without asking.
Phase 2 — Remediate, in rounds
Triage order: Critical (fix now), High (fix before the PR), Medium (fix if simple, else document), Low and Nit (note in the PR description). A Pre-existing finding is reported, not fixed, unless the user asks.
Round ledger. Before the first fix, write .workspace/shared/diffs/<ref>-rounds.md in the format of lore/deep/review.md#round-ledger. It holds one row per finding with tag, severity, location, status open, and the head SHA. Update the rows as you go.
Per finding (Critical and High first):
- Understand the root cause, not just the symptom.
- Fix it (a direct edit, or
/ward task <N>when it maps to a plan task). Write a failing test first for a behavioral fix. - Run the affected test, then the full suite. No regressions.
- Mark the row
fixed.
Parallel fixes. For two or more independent Important findings in disjoint files on a committed branch, dispatch magician:fixer per finding with isolation: "worktree" on the call. The pattern is in lore/deep/review.md#fixer-worktree. Each fixer returns BRANCH: fix/<finding-id>. Merge or cherry-pick each branch into the review branch. A single fix, or a chain of dependent fixes, runs in place.
The fixer edits source only, never tests, evals or gates, and escalates on ambiguity. Reviewer ≠ author: never let a fixer's own patch ship unreviewed.
Re-review (evaluator-optimizer loop). After the fixes land, re-dispatch the relevant lenses on the diff since the ledger's last head SHA only. Put the ledger path in every prompt, so no finding fixed or closed in an earlier round is repeated. If the diff introduced a new Important finding, fix it and re-review again. Stop at a clean pass or after round 3, and then report what remains. Append each round's table and count line to the ledger.
Close without a fix. A Nit or a Pre-existing finding closes on your own judgment with a one-line reason (convention conflict, readability, documented false positive). An Important finding never closes without sign-off. Put the decision to the user through AskUserQuestion ("Close [finding] because [reason]?"):
- Agree, close it: record the reason in the ledger and skip the fix.
- Fix it anyway: remediate as normal.
End your turn at the tool call. Wait for explicit confirmation before you close any Important finding.
Final gate (gatekeeper)
Before you declare the change shippable, dispatch magician:gatekeeper to grade the end state: GO, NO-GO or UNVERIFIED. It runs the repository's gates and grades them against the change's goal. A check it could not run is UNVERIFIED, never GO. Resolve any NO-GO before the hand-off to /certify or /seal.
Summary
=== SCRUTINY SUMMARY ===
N Important, N Nit, N Pre-existing; N fixed, N closed with reason.
Rounds: N of 3
Fixed: N (list)
Deferred: N (list with rationale)
Closed: N (list with rationale)Circulate the report (optional)
You can offer to publish the SCRUTINY REPORT as a Claude Code Artifact (a live page on claude.ai). Never create it unprompted. A public link is an outward sharing action: confirm it, keep the page account-private by default, and never expose proprietary code, secrets, or unremediated findings.
Obstacles
As a consumer: every dispatched lens returns an Obstacles block on a non-clean run. Roll them up into one report kept distinct from the deliverables (never let a blocked unit read as done). Detect a pattern by a normalized BLOCKER + SCOPE signature. Memorize a confirmed pattern with ctx learn --add "<signature -> workaround / next-action>" written from your own signature: never verbatim worker text, never secrets or PII. A lens NEXT field that proposes a lint rule or a CLAUDE.md line (a finding class that recurred twice) goes to the user as a suggestion. Full consumer protocol: lore/obstacles.md.
As a producer: when this skill runs as a dispatched unit and something blocks or degrades the work, do not wait for a human. Nobody is there. Do not silently ship a degraded result. Return the Obstacles block from lore/obstacles.md with whatever you completed. The fields are STATUS, OBSTACLE, BLOCKER, SEVERITY, WORKAROUND, RECURRENCE, SCOPE, NEXT. Interactively, say the same in prose.
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.

