Generate a CLAUDE.md your agent will actually follow

Most instruction files are ignored because they are bloated or vague. Answer 7 questions and get a short, tailored CLAUDE.md or AGENTS.md that tells your agent how to behave: ask first, stay in scope, and not ship slop. Free and instant.

What are you building?

Package manager or runtime?

How do you check your work?

How much should the agent ask before building?

Solo or team?

How strict on scope?

Shipping user-facing UI or copy?

Which file do you want?

Both is the officially documented pattern: AGENTS.md holds the rules, and a tiny CLAUDE.md imports it, so one file works across Claude Code, Cursor, and other agents.

AGENTS.md

# Agent guide

This is a Next.js (App Router) project. Use npm for scripts. Keep this file short: only rules you would actually enforce, nothing the code already makes obvious.

## Commands
- Dev: `npm run dev`
- Test: `npm test`
- Lint: `npm run lint`

## Before you build
Before writing any non-trivial code, ask me 3 to 4 clarifying questions as multiple choice (A/B/C, and mark the option you recommend). Then restate the goal in one line and wait for my yes. Do not start building on assumptions.

## Rules
- NEVER add features, dependencies, refactors, or files I did not ask for. If you think something else is needed, propose it first and wait.
- Make the smallest change that solves the task. Do not rewrite a whole file for a one line fix.
- NEVER guess a framework API from memory. This project's Next.js version may differ from your training data. Check the installed version's behavior before you write against it.
- Prefer Server Components. Add "use client" only where you genuinely need interactivity.
- Commit only when I ask you to.

## Off-limits
- Do not edit lockfiles, generated output, or build directories.
- Never put a secret in client-side code, and never commit a .env file.

## Definition of done
There is no test suite yet, so verify your change in the running app and tell me exactly what you checked. Never claim a task is done without confirming it actually runs.

## Writing copy and UI
- No em-dashes. No buzzwords (delve, seamless, robust, elevate, unleash). No walls of emoji headings.
- Write like a real person. Match the existing design system; do not invent new colors, fonts, or flat gradients.

## Keeping this file honest
Add a rule only after the agent makes the same mistake twice. If a line is not one you would enforce, delete it. Aim to stay under about 80 lines.

This file asks nicely. It cannot enforce anything. kitstarter makes the ask-first rule actually stick with hooks and a live tutor, so your agent follows it instead of drifting.

See how the kit enforces it

This is a free CLAUDE.md and AGENTS.md generator. You answer a few multiple-choice questions about your stack, how much you want the agent to ask before building, and how strict to be about scope, and it produces a short instruction file (about 40 to 80 lines) tuned to those answers. Unlike other generators it encodes behavior, an ask-first gate, scope discipline, and an anti-slop rule, not just your commands. It runs entirely in your browser, so it is instant and free.

Why this one is different

Every other generator writes down your stack. This one writes down how the agent should behave.

It generates behavior, not just facts

Other generators emit your stack and your commands. This one adds the rules that change how the agent acts: when to ask, what not to touch, and how to know it is done. That is the actual reason agents go off the rails.

Built to be followed, not ignored

Short by design, with MUST and NEVER wording and a do-this-instead for every rule. Long, vague instruction files get skipped; a tight one gets obeyed. We keep the output under about 80 lines on purpose.

An honest bridge to the real fix

A file can only ask nicely; it cannot enforce anything. That is the whole point of the kit: hooks and a tutor that make the ask-first rule actually stick. The generator gives you the file free, and shows you where it stops.

Common questions

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

CLAUDE.md is the instruction file Claude Code reads. AGENTS.md is the cross-tool open standard that Cursor, Codex, and others also read. The recommended setup is a single AGENTS.md with the rules and a short CLAUDE.md that imports it, so you maintain one file. This generator can emit either or both.

Why do agents ignore CLAUDE.md?

Because the file is loaded as advice, not enforced, and because bloated or vague files train the agent to skip them. The fixes are to keep it short, use direct MUST and NEVER wording, pair every prohibition with what to do instead, and never restate what a linter already checks. This generator follows all of that.

Is it really free, and does it send my code anywhere?

It is free and it runs in your browser. It does not read your code or your repository. It only uses the answers you pick, so there is nothing to upload. If you choose to have the file emailed to you, that is the only time an email address is involved.

Will this stop my agent from over-building or ignoring instructions?

It helps: a short, direct, behavior-first file is followed far more often than a vague one. But an instruction file can only ask. To actually enforce ask-first and catch mistakes as they happen, you need hooks, which is what the kitstarter kit adds on top of the file.

Make your agent ask before it builds

One kit. Ask-first behavior, lean output, no AI slop. Works with Claude Code, Codex, and Antigravity.

Get the kit · $20 $29

One-time. Less than one hour of cleaning up AI slop.