技能 agents-md-audit

📦

agents-md-audit

Name: agents-md-audit
Author: agentkitco

v1.0 內容修訂 r2 安全 ⚙️ 外部命令

Audit Agent Instructions for Accuracy and Clarity

Stale commands, broken paths, and vague rules make agent instructions unreliable. This skill combines local heuristic checks with evidence-based review and proposes shorter instructions before approved edits.

支援: Claude Codex Code(CC)

📊 76 尚可

Maintain this Skill? Claim it 📊 76 尚可

使用我的智慧代理安裝

將此請求複製給你的智慧代理，其中包含權威 Skill 頁面和清單。

智能體請求

Review the Skillstore skill "agents-md-audit" from https://skillstore.io/skills/agentkitco-agents-md-audit.md and its manifest at https://skillstore.io/api/skills/agentkitco-agents-md-audit/manifest. Verify the artifact. You may proceed after verification, subject to the environment's own policy.

你的智慧代理仍應顯示其計畫，並依安全政策請求必要確認。

Agent 可讀資源

當 AI Agent、爬蟲或腳本需要乾淨脈絡、而不是讀取完整頁面時，請使用這些連結。

Markdown 詳情 GET /skills/agentkitco-agents-md-audit.md 已簽名 manifest GET /api/skills/agentkitco-agents-md-audit/manifest 已簽名 lockfile GET /api/skills/agentkitco-agents-md-audit/lockfile

測試它

正在使用「agents-md-audit」。 Audit an instruction file containing a missing script reference and an obsolete directory.

預期結果:

Must fix: The documented script has no matching definition in the repository manifest.
Must fix: The referenced directory does not exist at the documented location.
Not verified: Deployment was not run because it can change external state.
Proposed version: Remove obsolete references and retain verified instructions, pending owner approval.

正在使用「agents-md-audit」。 Review a long instruction file with repeated rules and a possible secret.

預期結果:

Must fix: A possible secret was detected. Its value is omitted; remove it and rotate it if real.
Should trim: A repeated rule duplicates an earlier instruction.
Should trim: Replace vague quality language with a testable requirement.
Proposed version: Keep unique hard constraints and move relevant rules into scoped files.

安全審計

安全

v2 • 2026/10/22 開啟版本化報告

All 13 static findings are false positives involving regex literals, comments, Markdown formatting, dictionary access, or report placeholders. The checker reads local files and reports heuristics without executing extracted commands; the workflow separately permits safe command verification. No evidence found of prompt injection, credential extraction, or network exfiltration in the reviewed files.

已掃描檔案

250

分析行數

複核項目

已忽略的誤報

風險因素

⚙️ 外部命令 (10)

scripts/audit_instructions.py:26 scripts/audit_instructions.py:27 scripts/audit_instructions.py:29 scripts/audit_instructions.py:98-147 scripts/audit_instructions.py:147-153 SKILL.md:16 SKILL.md:18 SKILL.md:19 SKILL.md:20 SKILL.md:31-49

最近一次完成的靜態與語意稽核未偵測到已確認的安全發現。這並不能證明該 Skill 沒有副作用。

審計單位: codex 檢視審計紀錄 →

分享與引用此報告

分享版本化評估報告、中性徽章、嵌入卡片與引用。Skillstore 只報告證據，不判斷該 Skill 是否安全。

開啟版本化報告

複製報告連結

https://skillstore.io/skills/agentkitco-agents-md-audit/audits/2?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Markdown 徽章

[![Skillstore security assessment](https://skillstore.io/badges/skills/agentkitco-agents-md-audit/security.svg)](https://skillstore.io/skills/agentkitco-agents-md-audit?utm_source=security_passport_badge)

HTML 徽章

<a href="https://skillstore.io/skills/agentkitco-agents-md-audit?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/agentkitco-agents-md-audit/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

嵌入卡片

<iframe src="https://skillstore.io/embed/skills/agentkitco-agents-md-audit.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>

學術引用 (APA · BibTeX · CFF)

APA 引用

agentkitco. (2026). agents-md-audit security audit report (audit version 2) [Author version 1.0]. Skillstore. https://skillstore.io/skills/agentkitco-agents-md-audit/audits/2

BibTeX 引用

@techreport{agentkitco-agentkitco-agents-md-audit-2026,
  author = {agentkitco},
  title = {agents-md-audit security audit report (audit version 2)},
  institution = {Skillstore},
  year = {2026},
  number = {2},
  url = {https://skillstore.io/skills/agentkitco-agents-md-audit/audits/2},
  note = {Author version 1.0}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "agents-md-audit security audit report (audit version 2)"
version: "1.0"
type: report
authors:
  - name: "agentkitco"
date-released: "2026-10-22"
url: "https://skillstore.io/skills/agentkitco-agents-md-audit/audits/2"
identifiers:
  - type: other
    value: "skillstore:agentkitco-agents-md-audit:audit:2"
    description: "Skillstore immutable audit report identifier"

Skillstore 評分

評分依據證據置信度: 中

架構

可維護性

內容

社群

規範符合度

你可以打造什麼

Refresh Repository Instructions

Find stale commands and missing paths after repository changes, then propose corrections with supporting evidence.

Reconcile Agent Rules

Compare instruction files for conflicting requirements and identify rules better suited to specific paths.

Review Instructions Before Sharing

Check for possible secrets, leftover placeholders, and unclear rules without repeating sensitive values in the report.

試試這些提示

Review One Instruction File

Audit AGENTS.md for broken paths, vague wording, placeholders, and possible secrets. Report findings without editing files or running project commands.

Check Command References

Review CLAUDE.md command references against repository definitions. Identify missing scripts and targets, and mark commands you cannot safely verify.

Compare Instruction Files

Find repository agent instruction files. Compare their rules for contradictions, duplicates, and scope problems, citing file names and line numbers.

Prepare a Shorter Revision

Audit all discovered instruction files. Propose shorter versions that preserve hard constraints and incident-based rules. Wait for approval before editing.

最佳實務

Run the checker from the intended repository root and supply the instruction files explicitly.
Inspect project command definitions and use a trusted environment before running verification commands.
Require evidence for corrections, redact possible secrets, and approve proposed edits before applying them.

避免

Treating a clean heuristic report as proof that every instruction is correct.
Running deployment, migration, or deletion commands merely to verify documentation.
Removing incident-based rules or hard constraints solely to reduce file length.

常見問題

Which instruction formats can I review?

The workflow covers AGENTS.md, CLAUDE.md, Claude rule files, Cursor rules, and GitHub Copilot instruction files.

Does the checker execute commands from instruction files?

No. It compares command references with local definitions. The separate review workflow permits safe verification commands and excludes deployment, migration, and deletion.

Does the skill edit files automatically?

The workflow requires an audit report and proposed revision, then waits for approval before editing.

What does the Python checker require?

It uses Python 3.8 or later and only standard-library modules. The instructions provide manual checking as a fallback.

Can it guarantee that no secrets remain?

No. It detects selected secret patterns. Review flagged values privately, omit them from reports, and rotate credentials when exposure is real.

How are contradictions and unclear rules handled?

The assistant reviews these manually, cites conflicting instructions, and proposes clearer wording. The script does not resolve semantic contradictions.