Coding agent instructions
Generate a concise repository instruction file for Codex, Claude Code, Cursor, and other coding agents. Keep the rules actionable, testable, and small enough that they do not waste context.
Good agent instructions describe the repository, the commands that must run, the areas that should not be touched, and the verification expectations. They should not repeat the whole README or become a long style guide that every task has to carry.
People searching for an AGENTS.md generator usually want one durable instruction file that keeps coding agents from repeating the same mistakes. The safest pattern is to make AGENTS.md the concise operational source of truth, then bridge tool-specific files to it when needed. That avoids maintaining three long, drifting instruction documents.
Codex has first-class AGENTS.md discovery. OpenAI's Codex documentation says it reads AGENTS.md files before work starts, layers global and project guidance, and lets closer files override broader ones. Claude Code uses memory files such as CLAUDE.md, and its documentation recommends using `/memory` to verify what loaded. Cursor uses project rules, commonly stored under `.cursor/rules`, and the older `.cursorrules` format has been treated as a legacy path in Cursor docs.
| Tool | Primary guidance file | Best use | Risk to avoid |
|---|---|---|---|
| Codex | AGENTS.md and optional nested overrides. |
Repository commands, test expectations, deployment boundaries, and local conventions. | Overly large files that hit instruction-size limits or conflict with deeper directory rules. |
| Claude Code | CLAUDE.md, often importing or summarizing AGENTS.md. |
Small project memory, recurring corrections, model/workflow hints. | Long stale memory files that consume context on every turn. |
| Cursor | .cursor/rules/*.mdc or project rules. |
Scoped rules for frameworks, directories, test commands, and coding standards. | A broad always-on rule set that makes every task carry irrelevant context. |
A useful AGENTS.md should read like an internal runbook for an agent that can edit files. It should not be a marketing README, a full architecture book, or a copy of framework documentation. The generated template uses six sections because they map to the moments where agents need concrete direction: project context, commands, coding rules, safety/data rules, verification, and final response expectations.
For a static website or pSEO site, AGENTS.md should mention canonical URL rules, sitemap updates, analytics/ad placement constraints, and the public-only deploy directory. For a React or Next.js app, it should include package manager, build command, component conventions, route structure, and accessibility checks. For a backend, the file should make database migration, auth, secrets, and test fixture rules explicit. For a monorepo, keep the root file short and add path-scoped guidance near specialized packages.
The generator intentionally creates a compact file. If the output grows beyond a few hundred lines, split it. A root AGENTS.md can hold global rules, while `apps/web/AGENTS.md`, `services/billing/AGENTS.md`, or `.cursor/rules/frontend.mdc` can hold scoped rules. Smaller scoped rules are easier to keep current and less likely to pollute unrelated tasks.
For project-specific instructions, yes. Versioning the file keeps agent behavior aligned with code changes and lets teammates review changes to commands, safety rules, and verification policy.
It can contain operational rules, but not secrets. Do not include API keys, cookies, account recovery details, private customer data, or anything that would be harmful if committed or accidentally deployed.
Short enough that an agent can use it on every task. A practical root file is often 50-150 lines. If a rule only applies to one directory or framework, put it in a scoped file instead of the root.
This generator is independent. Confirm exact file loading behavior in the coding agent you use because tools can differ in precedence and scope rules.