- Home
- Skills
- Prompt & Skill Engineering
- AI Agent Instructions Check
More screenshots
Works with the AI Agents you already use
AI Agent Instructions Check
This skill finds those and seven other mistakes in agent instruction files, and tells you how to fix each one.
$5
AI Agent Instructions Check
Example session with this skill installed
Check this project with ai-agent-instructions-check.
- Read your context and instructions
- Compiled the ai agent instructions
$ python3 scripts/check_agent_instructions.py examples/risky
BLOCK AGENTS.md:13 A02 SAFETY_CHECKS_TURNED_OFF
- Commit with git commit --no-verify so the slow pre-commit hook does not block you.
Why: The file tells the agent to switch off a safeguard: commit or push with --no-verify (skips git hooks), run without permission prompts or sandbox (--dangerously-skip-permissions, bypassPermissions, --yolo, danger-full-access, --allow-all-tools), force-push to main, rm -rf on /, ~, * or ., delete failing tests, disable hooks or branch protection, or never ask before deleting, pushing or deploying. The agent applies it to every task, including the one where the safeguard would have caught a mistake.
Safer: Remove it or write it as a prohibition ("Never use --no-verify"). If a step really needs it, say when and require the user's explicit approval; run unattended agents only in a container or sandbox.
BLOCK AGENTS.md:14 A02 SAFETY_CHECKS_TURNED_OFF
BLOCK AGENTS.md:16 A03 HIDDEN_INSTRUCTIONS
BLOCK CLAUDE.md:8 A01 SECRET_IN_INSTRUCTIONS
BLOCK CLAUDE.md:13 A01 SECRET_IN_INSTRUCTIONS
BLOCK CLAUDE.md:17 A02 SAFETY_CHECKS_TURNED_OFF
BLOCK CLAUDE.md:18 A04 DECEPTIVE_INSTRUCTION
REVIEW .claude/rules/api.md:2 A13 RULE_FRONTMATTER_PROBLEM
REVIEW .cursor/rules/style.md:1 A13 RULE_FRONTMATTER_PROBLEM
REVIEW .cursor/rules/testing.mdc:3 A13 RULE_FRONTMATTER_PROBLEM
REVIEW .github/copilot-instructions.md:5 A12 MISSING_REFERENCED_FILE
REVIEW AGENTS.md:7 A15 PACKAGE_MANAGER_MISMATCH
REVIEW AGENTS.md:8 A14 DOWNLOAD_AND_RUN
REVIEW AGENTS.md:9 A16 PERSONAL_PATH
REVIEW CLAUDE.md:7 A12 MISSING_REFERENCED_FILE
REVIEW packages/api/CLAUDE.md:201 A11 FILE_TOO_LONG
(Each finding also has Why and Safer lines, as in the first one.)
Verdict: BLOCK. Checked 7 files, 7 with agent instruction content.
This is a static check of file text. NOT FLAGGED is not a guarantee.
Connects securely to your tools. The creator never sees your data.
About this skill
One line in AGENTS.md such as "commit with --no-verify" turns off your git hooks for every task the agent does, a token pasted into CLAUDE.md is sent to the model provider with every request, and an HTML comment that GitHub never shows can tell the agent to run a script. This skill finds those and seven other mistakes in agent instruction files, and tells you how to fix each one.
Your agent runs a scanner on AGENTS.md, CLAUDE.md and the .claude rules, subagents and slash commands, GEMINI.md, Cursor .mdc rules and .cursorrules, Copilot's copilot-instructions.md and *.instructions.md, and Windsurf and Cline rules, in every folder of the project. It tells prohibitions from orders ("Never use --no-verify" is fine), treats HTML comments and invisible Unicode as hidden text, reads @path imports the way Claude Code does (relative to the file, not inside code), and checks each tool's frontmatter keys against that tool's documentation. It never runs anything the files say. It returns BLOCK, REVIEW, or NOT FLAGGED, with the file, line, reason, and a safer approach. Secrets are never printed.
What it checks
BLOCK
- A01: passwords, tokens and API keys in the instructions, or @.env and key files imported into them
- A02: orders that switch off safeguards: --no-verify, --dangerously-skip-permissions, --yolo, force-push to main, rm -rf ~, never ask before deleting or deploying, delete failing tests
- A03: hidden text: HTML comments that carry commands or deception, Unicode tag characters, bidi overrides and zero-width runs
- A04: orders to hide actions from the user, override the user's or system instructions, or send secrets to an outside address
REVIEW
- A11: files longer than the tool reads well (CLAUDE.md over 200 lines, .mdc rules over 500, AGENTS.md over 32 KiB)
- A12: @path imports, mdc: links and Markdown links to files that do not exist
- A13: rule frontmatter that does not work: .md files in .cursor/rules, alwaysApply with globs, globs instead of paths, unquoted * patterns
- A14: downloads piped into a shell (curl | sh, iwr | iex)
- A15: install commands for a package manager the project does not use
- A16: personal home-folder paths such as C:\Users\name or /Users/name
What's included
- The scanner (Python 3.9+, standard library only)
- A rules guide with the reason and safer approach for each rule
- A risky and a safe example project, and the real report the scanner produces on the risky one
Limits It reads each file as written: skills in .claude/skills, user-level files outside the project (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, global Cursor or Windsurf rules), settings that change which files load, and the content of imported files that are not instruction files are not checked. Whether a sentence is an order or a prohibition is decided from its wording, so unusual phrasing can be missed. NOT FLAGGED is not a guarantee.
Static check of instruction file text. User-level and global rule files, settings that change which files load, and unusually worded sentences are not fully covered. NOT FLAGGED is not a guarantee.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 5 days ago
Requires
Permissions
Runs one local Python script in the terminal. It reads your project's files, writes nothing, and makes no network calls. The example project includes a clearly fake, inert token (not-a-real-...) and a labeled sample hidden comment for the scanner to find. Nothing in it is run.