Free tool · runs in your browser

Coding agent rules generator

Get a short rules file for Claude Code, Cursor, Codex, GitHub Copilot or any coding agent, in the format and folder that agent reads. It comes with a project spec, a build plan with checks for every phase, and the first prompt to paste. The built-in check scores the rules file, or one you already have.

promptup — agent rules generatorNothing leaves your device

What are you building?

Project type

Coding agent

Your experience

Jump to your kit

Stack

Only choices that work together are shown. Each one says what you trade.

Features for version 1

How the agent should work

Rules only you know

These matter most, because an agent cannot guess them. The rules check flags vague ones.

More details (optional)

Project name Who uses it Not in version 1 (one per line)

CLAUDE.mdRules 100

# Booking app for yoga studios

A booking app for yoga studios where members reserve classes and pay a monthly plan. Stack: Next.js (TypeScript), Postgres with Prisma, Better Auth, Stripe, Tailwind CSS, Vercel. Spec: SPEC.md. Build plan: PLAN.md.

## Commands
- Install: `npm install`
- Dev server: `npm run dev`
- Tests: `npm test` (Vitest)
- End-to-end tests: `npm run test:e2e` (Playwright)
- Lint: `npm run lint`
- Type check: `npm run typecheck`
- Build: `npm run build`
These are the planned commands. If setup changes any of them, update this list in the same commit.

## Stack and conventions
- Next.js App Router with TypeScript. Components are server components by default; add "use client" only where a component needs state, effects or browser APIs.
- Only env vars that start with `NEXT_PUBLIC_` reach the browser. Every other secret stays in server code.
- TypeScript strict mode. No `any`: use `unknown` and narrow it.
- Database schema lives in `prisma/schema.prisma`. After changing it, run `npx prisma migrate dev` with a short `--name`. Never edit an applied migration.
- Check the signed-in user on the server for every protected page, action and API route. Hiding a button is not access control.
- Payment status comes from Stripe webhooks: verify the signature, handle each event once even if it arrives twice, and never mark anything paid from the success-page redirect.
- Use Stripe test-mode keys outside production.
- Style with Tailwind classes. No new CSS files apart from the one global stylesheet.
- Hosting is Vercel. Functions are short-lived: no background loops and no writing to the local disk. Env vars are set in the Vercel dashboard.
- All email goes through one helper module. Outside production it goes to the console or a test inbox, never to real people.

## How to work
- Plan first when a change touches more than 2 files: list the files and steps, then wait for my OK.
- Work one phase of PLAN.md at a time and stop at the end of each phase with the check results.
- Add or update a test for every change in behaviour. A bug fix starts with a test that fails.
- Ask before adding a dependency, and say what it is for.
- Change only the files the task needs: no drive-by refactors, renames or reformatting.

## Never
- Commit `.env` files or print secrets. Add each new variable to `.env.example` with a dummy value.
- Run destructive commands without asking: deleting data, dropping or resetting the database, `git push --force`.
- Edit generated folders (`node_modules/`, `.next/`) or lockfiles by hand.

## Before you say a task is done
1. Run `npm run typecheck`, `npm run lint` and `npm test`. All must pass.
2. Report what changed, the commands you ran with their result, and anything you could not check.

Get a second opinion on the spec and plan

Open in ChatGPT Open in Claude

Rules check

How to use it

Say what you are building in one or two sentences. Working on an existing project? Switch to "Existing project" and paste your package.json, pyproject.toml or file tree. Your real commands and tools are read in the browser.
Pick the project type and your coding agent. The rules file is written in that agent’s format, with the file name and folder it reads.
Choose the stack. Only options that work together are shown, each with a one-line trade-off, and clashing choices get a warning.
Add the rules only you know, such as how prices or dates are stored. Then check the score and fix anything the rules check flags.
Download the kit into your project folder, open it in your agent and paste the first prompt. The agent asks its questions and plans before writing code, then builds one phase at a time.

Which file each coding agent reads

Each agent looks for its instructions in a different place. The kit writes the one your agent reads, in its format.

Agent	File the kit writes	Format	How long
Claude Code	`CLAUDE.md`	Plain Markdown. It can pull in other files with `@path`.	Under 200 lines
Cursor	`.cursor/rules/project.mdc`	Markdown with a `---` header: `description`, `globs`, `alwaysApply`. Cursor also reads `AGENTS.md`.	Under 500 lines
Codex	`AGENTS.md`	Plain Markdown. A file in a subfolder adds to the one in the root.	32 KiB in total
GitHub Copilot	`.github/copilot-instructions.md`	Plain Markdown. Rules for some folders go in `.instructions.md` files with `applyTo`.	Code review reads the first 4,000 characters
Antigravity	`AGENTS.md`	Plain Markdown; it also reads `GEMINI.md`. Files in `.agents/rules/` need a `trigger` header.	24,000 bytes per rule file
Devin Desktop (formerly Windsurf)	`.devin/rules/project.md`	Markdown with `trigger: always_on`, `model_decision`, `glob` or `manual`. It also reads `AGENTS.md`.	12,000 characters per file
Gemini CLI	`GEMINI.md`	Plain Markdown, with `@file` imports.	No limit given
Several agents	`AGENTS.md`	One shared file. Claude Code reads it when there is no `CLAUDE.md`.	Keep it short

Checked against each tool’s documentation on 10 October 2026: Claude Code, Cursor, AGENTS.md, GitHub Copilot, Antigravity, Devin Desktop and Gemini CLI. For Google AI Pro, Ultra and free personal accounts, Gemini CLI was replaced by Antigravity CLI on 18 June 2026.

What makes a rules file work

Commands first. How to install, run, test and lint, each in backticks. An agent can read your code, but it cannot guess these.
Only what the agent would get wrong. Claude Code’s guide suggests asking of every line: would removing it cause mistakes? "Write clean code" never passes that test. "Prices are stored in paise" does.
A check before "done". Name the tests, lint and type checks that must pass. Without a check it can run, an agent stops when the work only looks finished.
One rule per line, and no contradictions. When two rules disagree, the agent may follow either one.
Short. A 2026 study of AGENTS.md files (ETH Zurich, arXiv 2602.11988) found that context files raised the cost of agent runs by over 20% and that repository overviews did not help. Its advice: describe only the minimal requirements.
The task lives elsewhere. Lasting rules go in the rules file. What you are building right now goes in SPEC.md and PLAN.md, so the rules file stays the same from one task to the next.

What the rules check looks for

Check	Why it matters
Length against the agent’s own limit	Long files make agents skip rules, and some agents cut a file off after a set size.
Commands near the top, including a test or check command	They are what the agent cannot work out, and a check lets it see when the work is really done.
Vague words such as "best practices", "properly" or "robust"	Nobody can tell whether a vague rule was followed, the agent included.
Things the agent does anyway, such as "clean code" or "meaningful names"	They take up space and push the rules that matter further down.
The same rule twice	Repeating a rule does not make it stronger. It only makes the file longer.
Conflicts: tabs and spaces, npm and pnpm, "always" and "never"	The agent picks one side at random.
Secrets: API keys, tokens, passwords	Rules files are committed to git and shared with every session.
Headers Cursor, Devin Desktop and Antigravity need	Without the right header, a rule may never apply. Antigravity drops it without telling you.
Too many IMPORTANT and MUST words	When every line shouts, none stands out.
Notes that go stale, such as "currently" or a date	Short-lived notes belong in the task, not in rules every session reads.

The score is a guide to what to fix, not a promise about the agent’s results. To tighten a single prompt rather than a rules file, use the Prompt Analyzer.

More coding prompt help

FAQ

Questions, answered

What is the difference between CLAUDE.md and AGENTS.md?

CLAUDE.md is the file Claude Code reads at the start of every session. AGENTS.md is an open format read by Codex, Cursor, GitHub Copilot, Devin Desktop, Antigravity and other agents. Claude Code also reads AGENTS.md when a project has no CLAUDE.md, so one AGENTS.md can serve every agent your team uses.

Where do I put the rules file?

CLAUDE.md, AGENTS.md and GEMINI.md go in the root of your project. Cursor rules go in .cursor/rules/, Copilot instructions in .github/, and Devin Desktop rules in .devin/rules/. The zip keeps these folders, so unzip it into your project folder.

How long should a rules file be?

As short as it can be while still holding everything the agent would otherwise get wrong. Claude Code’s docs say to keep CLAUDE.md under 200 lines, and Cursor’s say to keep rules under 500 lines. The files this tool writes are 28 to 44 lines before you add your own rules. The rules check shows your line count against your agent’s limit.

Is my code or package.json uploaded?

No. Everything happens in your browser, including reading a pasted package.json or file tree. Nothing you type or paste is sent to PromptUp or anyone else. Share links hold only your choices, never what you typed.

Build your coding agent kit

Rules check

Rules check

How to use it

Which file each coding agent reads

What makes a rules file work

What the rules check looks for

More coding prompt help

Questions, answered