Tools / AI Agent Guardrails
AI Coding Agent Guardrails Generator
Turn the rules your team already agrees on into guardrails an agent can follow: a rules file (AGENTS.md or CLAUDE.md), permission allow/ask/deny lists, protected paths and a pre-merge review checklist. Everything is generated locally in your browser.
The generated documents follow the page language: switch to 繁體中文 above if your team writes its internal docs in Chinese.
Check the key names and the file location against the documentation for the exact agent CLI version you run — configuration schemas change between releases.
What are agent guardrails?
Guardrails are the parts of a coding workflow you do not want to re-explain in every session: which commands verify a change, which files are off limits, and what the agent must never do on its own. They belong in version control next to the code, not in a chat thread that disappears.
Why a rules file beats a better prompt
- A prompt is written once and forgotten; a rules file is read by every session, by every teammate and by every agent you switch to.
- Rules that live in the repository can be reviewed like code, so the team can argue about them once instead of every time an agent surprises someone.
- Explicit "never touch this" lists turn a policy argument into a mechanical check: secrets, deployment configs and lockfiles stop being judgement calls.
The three permission lists
Allow what is read-only and what your project already runs (test, lint, build). Put anything that rewrites history, installs packages or reaches the network behind an explicit ask. Deny secrets, production configuration, deployment pipelines and destructive shell commands outright. Anything not listed should fall back to asking a human.
FAQ
Will this stop an agent from writing bad code?
No. Guardrails reduce the blast radius of a mistake, they do not replace review. The rules file makes the agent's constraints explicit, and the review checklist covers what a guardrail cannot: a change that is in scope, passes tests and still makes the design worse.
Where do the generated files go?
The rules file belongs in the repository root so it is committed and reviewed with the code. Permission settings usually live in a per-user or per-project settings file outside version control — that is deliberate, since they encode what you personally allow on your machine.
We use several agents. Do we need several rule files?
Keep one source of truth and copy it. The generated content is plain Markdown with no vendor-specific syntax, so the same text works as AGENTS.md, CLAUDE.md or any other instruction file an agent reads from the project root.