agentleFS
oldmilkyoldmilky/luno-for-cc/.claude/skills/audit/SKILL.md

audit

Skill0 starsChanged 4 months agoMITView on GitHub

Deep read-only review of the branch diff against a base ref, across several lenses, with every HIGH finding independently refuted before it is reported. Use when the user says /audit or wants a thorough review of a branch rather than the working tree.

193 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use delight, teach-impeccable and 10 more.

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.
  • HK-hubSkill

    delight

    Add moments of joy, personality, and unexpected touches that make interfaces memorable and enjoyable to use. Elevates functional to delightful. Use when the user asks to add polish, personality, animations, micro-interactions, delight, or make an interface feel fun or memorable.

    Also in 32 projects that use this

  • HK-hubSkill

    teach-impeccable

    One-time setup that gathers design context for your project and saves it to your AI config file. Run once to establish persistent design guidelines.

    Also in 19 projects that use this

  • zwy123zwySkill

    normalize

    将用户随意输入的提示词规范化为结构清晰、便于大模型执行的标准化提示词。适用于需要整理、统一提示词格式时调用。

    Also in 19 projects that use this

  • SkillPilotCHSkill

    level-up

    Eine Automatisierung pro Woche finden, scopen, shippen. Output landet als proposed-Eintrag in cadence.md mit konkreten Implementations-Schritten.

    Also in 12 projects that use this

  • ww-w-aiSkill

    pm-discovery

    PM Agent Team - Automated product discovery, strategy, and PRD generation. Runs PM analysis workflow to produce comprehensive PRD before PDCA Plan phase.

    Also in 3 projects that use this

  • +7 more from HK-hub

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 5 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 174 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every HIGH and CRITICAL goes to a skeptic whose brief is adversarial: _read the code and show this cannot happen_.

    line 53
  2. Distinctive:

    Default to refuted when uncertain.

    line 58
  3. Distinctive:

    Ask whether to fix the confirmed findings.

    line 92
  4. Distinctive:

    Read-only: No edits, no commits, no pushes.

    line 98
  5. Distinctive:

    Verification gates the report: Only what survived a skeptic is reported as confirmed.

    line 99

“Common” is measured: kept by at least 6 in 10 of the 174 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

A deep parallel read of the current branch's diff. One reviewer per lens, then skeptics who try to refute each finding, so plausible-but-wrong ones do not survive into the report.

$ARGUMENTS: first token is the base ref (default main), second is effort — medium (default) runs one skeptic per finding, high runs three and takes the majority.

This is the deeper sibling of /check. /check reads the working tree in seconds; /audit reads everything since a base ref and is worth the wait before something lands.

Phase 0 — Scope

git diff --name-only <base>...HEAD
git diff --stat <base>...HEAD

Drop lockfiles and pure-formatting churn — a reformat commit is not review material and will drown the signal. Empty list → say so and stop.

Note the shape early: a diff that is 90% one mechanical rename wants a different read than one that rewrites the permission path.

Phase 1 — Lenses, in parallel

Dispatch in one message so they actually run concurrently. Give each the file list and the base ref.

LensWhoLooks for
Securityextension-security-reviewerspawn, credentials, the permission gate, CSP, postMessage trust
Webviewwebview-design-reviewertokens, motion presets, icons, the seven traps
Correctnessgeneral-purposewrong conditions, unawaited promises, swallowed errors, stale state, races
Contractgeneral-purposethe postMessage union — a message added on one side and unhandled on the other, a field the host reads but the webview never sends, a type widened to make it compile

The contract lens is specific to this codebase and earns its place: the protocol is the seam between two separately-compiled programs, so nothing enforces it but review. An unhandled message type fails silently at runtime.

Skip a lens whose files are untouched, and say which you skipped.

Phase 2 — Refute

Every HIGH and CRITICAL goes to a skeptic whose brief is adversarial: read the code and show this cannot happen. It must produce a concrete input, sequence or state — "looks fine to me" is not a refutation, and neither is "looks bad" a confirmation.

Default to refuted when uncertain. A review that reports ten things of which six are real trains the reader to skim; one that reports four real ones gets read.

At high effort, three skeptics per finding with different angles — does it reproduce, is the input actually reachable, does an existing guard already cover it — and the majority decides.

Phase 3 — Report

## /audit — HEAD vs <base> (<n> files, effort <e>)

### Confirmed HIGH / CRITICAL (<n>)
1. **<title>** — [<file>:<line>](<file>#L<line>) · <severity>
   <what breaks>
   <the input or sequence that breaks it>
   Fix: <concrete>

### Confirmed MEDIUM / LOW (<n>)
- <title> — [<file>](<file>) — <one line>

### Refuted (<n>)
- <title> — <why it cannot happen, one line>

### Not reviewed
- <lens skipped and why · files dropped as churn>

### Verdict
clean · <n> confirmed · <n> HIGH+ — fix before this lands

Phase 4 — Offer, do not act

Ask whether to fix the confirmed findings. Nothing is edited during an audit. Fixing is a separate step the user opts into, so the review and the change never blur into one diff.

Hard rules

  • Read-only. No edits, no commits, no pushes.
  • Verification gates the report. Only what survived a skeptic is reported as confirmed.
  • No lint-level noise. eslint, stylelint and prettier already ran; repeating them is not a finding.
  • Say what you did not look at. A silent gap reads as a clean bill of health.
  • A claim about rendered behaviour needs /browser, not an assertion.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.