Documentation

Agent Maker generates instruction files that give AI coding assistants the context they need to work effectively in your codebase. This guide covers what each file format does, why you need one, and how to get the most out of the wizard.

Why Use Instruction Files?

AI coding assistants like Claude Code, Gemini CLI, and other agents read instruction files from your project root to understand how your codebase works. Without one, the agent has to guess your conventions, tech stack, and project structure every session.

With a well-written instruction file, the agent:

• Follows your conventions — formatting, naming, import style, and project patterns from the first interaction
• Knows your commands — install, build, test, lint without asking
• Understands your architecture — folder structure, key modules, and how they connect
• Avoids common mistakes — respects your "do not" rules and security boundaries
• Plans before coding — with the bootstrap prompt, the agent maps the codebase before making changes

File Formats

CLAUDE.md

Read by Claude Code(Anthropic's CLI) and Claude-based agents. Placed at your project root, it's automatically loaded into context at the start of every conversation. Claude Code specifically looks for this file and treats it as project-level instructions.

Best for teams using Claude Code as their primary AI assistant.

AGENTS.md

Follows the agents.md open standard — a vendor-neutral format designed to work across any AI coding agent. The structure focuses on build/test commands, code standards, and testing requirements in a way that any agent can parse.

Best for teams that use multiple AI tools or want a future-proof format.

GEMINI.md

Read by Gemini CLI and Gemini-based tools. Mirrors the CLAUDE.md structure with Gemini-specific phrasing and conventions.

Best for teams using Google's Gemini CLI.

All Three

Generates all three files at once from the same wizard inputs. Each file is tailored to its target tool while containing the same project information. This is the recommended option if your team uses more than one AI assistant.

The Wizard — Step by Step

Step 1: Choose Your Format

Select which file(s) to generate. You can change this at any point during the wizard using the format switcher in the footer — your inputs are preserved.

Step 2: Project Overview

Enter your project name and a description of what it does. The description should focus on purpose and scope — what problem the project solves, who it's for, and what makes it different.

AI Suggest:If you have an API key saved, click "AI Suggest" to generate or improve the description. If you've already written something, it will refine your text. If the field is empty, it generates a description from the project name.

Step 3: Tech Stack

Tag your languages, frameworks, package manager, and any other tools (databases, Docker, etc.). Click suggestions to add common options, or type custom values and press Enter.

AI Suggest: Generates a tech stack recommendation based on your project description.

Step 4: Code Style

Define your formatting preferences: formatter (Prettier, Biome, ESLint, or none), indentation (2 spaces, 4 spaces, or tabs), quote style (single or double), and whether you prefer named exports.

These get translated into concrete rules in the generated file so the agent writes code that matches your existing style.

Step 5: Commands

Enter the shell commands for install, dev, build, test, lint, and type-check. Defaults are pre-filled based on your package manager selection — change them if your project uses custom scripts.

Step 6: Architecture

Describe your folder structure and key architectural decisions. Paste a folder tree, describe patterns (monorepo, microservices, etc.), or note important conventions.

AI Suggest: Generates a folder structure based on your selected languages and frameworks.

The "Additional Notes" field is for anything else the agent should know — security rules, deployment notes, environment quirks, or team conventions.

Step 7: Review and Export

Preview the generated markdown, switch between file formats (if you chose "All Three"), and export via copy or download. You can also save to the dashboard for later editing.

Plan-First Bootstrap Prompt

The bootstrap toggle (on by default) injects a special section at the top of your generated file. When an AI agent reads the file for the first time, it's instructed to:

1. Enter plan mode and explore the full codebase
2. Analyze dependencies, patterns, and architecture
3. Write a detailed implementation plan
4. Replace the bootstrap section with the plan in the file itself
5. Commit the updated file so future sessions benefit

After the first run, the bootstrap prompt is gone and the file contains a permanent implementation plan. This means every future session starts with a deep understanding of your codebase — no re-exploration needed.

Turn this off if you prefer to write the plan yourself or if your project is too small to need one.

AI Suggestions (BYOK)

AI Suggest buttons throughout the wizard use your own Anthropic API key to call Claude. This is a Bring Your Own Key (BYOK) model:

• Save your key in Settings— it's encrypted with AES-256-GCM and stored in the database
• The key is only decrypted server-side when making API calls — it's never sent to your browser
• If you don't have a key, the suggest buttons are disabled with a tooltip — the wizard works fine without them

Saving and Loading

Click Save on the Review step to persist your wizard state to the dashboard. Loading a saved project restores all your inputs to exactly where you left off.

If you save a project that already exists in the dashboard, it updates the existing entry instead of creating a duplicate. The button label changes from "Save" to "Update" to reflect this.

Tips for Better Instruction Files

• Be specific about what NOT to do— agents learn as much from constraints as from instructions. "Do not use axios — use native fetch" is more useful than "use fetch."
• Include your folder structure — even a rough tree helps the agent navigate without exploring every directory
• Keep commands accurate — wrong build or test commands cause the agent to fail silently or retry in loops
• Update the file as your project evolves — an outdated instruction file is worse than none. Re-run the wizard or edit directly when your stack changes.
• Use the bootstrap prompt for large projects — the upfront planning cost pays for itself in every subsequent session