Initialize Repository for AI Assistants

Generate an

INIT.md

and

AGENTS.md

at the repo root by deeply analyzing the codebase. These files onboard future AI coding assistant instances so they can be productive immediately.

Phase 1: Discovery

Thoroughly explore the codebase before writing anything. Read files in parallel where possible.

1.1 Existing documentation

Read these files if they exist (skip silently if missing):

• README.md

• CLAUDE.md

• AGENTS.md

• INIT.md

• CONTRIBUTING.md

• docs/

  directory (scan for key files)

1.2 Existing AI/editor rules

Read these if they exist:

• .cursor/rules/

  (all files)

• .cursorrules

• .github/copilot-instructions.md

1.3 Build & config files

Read root-level config files to extract commands and tech stack:

• package.json

  ,

  package-lock.json

  or

  yarn.lock

  or

  pnpm-lock.yaml

• pyproject.toml

  ,

  setup.py

  ,

  setup.cfg

  ,

  requirements.txt

  ,

  Pipfile

• Cargo.toml

• go.mod

• Makefile

  ,

  Justfile

  ,

  Taskfile.yml

• docker-compose.yml

  ,

  Dockerfile

• nx.json

  ,

  turbo.json

  ,

  lerna.json

  (monorepo configs)
• CI configs:

  .github/workflows/

  ,

  .gitlab-ci.yml

  ,

  Jenkinsfile

• Linter/formatter configs:

  .eslintrc*

  ,

  .prettierrc*

  ,

  ruff.toml

  ,

  pyproject.toml [tool.*]

  ,

  rustfmt.toml

  ,

  .golangci.yml

1.4 Architecture mapping

Explore top-level source directories (e.g.,

src/

,

app/

,

backend/

,

frontend/

,

services/

,

packages/

,

lib/

,

cmd/

):

• Identify entry points (main files, index files, route definitions)
• Identify service boundaries, API layers, data models
• Map the data flow: where requests enter, how they are processed, where data is stored
• Note key abstractions and patterns (middleware, plugins, dependency injection, etc.)

Do NOT read every file. Focus on entry points, config, and structural files to build a mental model.

Phase 2: Generate INIT.md

Create

INIT.md

at the repository root. Follow these rules strictly.

Required preamble

The file MUST begin with this exact text block:

You can write your own and get the same result. I use a custom /initialize on new repos to get up and running fast. I tweaked it to work across different coding agents and sprinkled in a few tips I collected along the way.

It should create a relevant AGENTS.md; if one exists, it updates it. Save this prompt as a custom command and use it with any tool — Gemini CLI, Codex, Amp, Firebender, etc. You aren't stuck with any single tool.

One more tip: a reasoning model works best for these types of commands.

Follow the preamble with a horizontal rule (

---

), then the content sections below.

Section 1: Commands

Header:

## Commands

List the concrete commands needed to develop in this codebase. Extract these from config files, Makefiles, READMEs, and CI configs. Include:

• How to install dependencies
• How to build / compile
• How to run the dev server
• How to run the full test suite
• How to run a single test file or test case
• How to lint and format
• How to run type checks
• Any other common commands (migrations, codegen, Docker, etc.)

Only include commands that actually exist in the project. Do not invent commands.

Section 2: Architecture

Header:

## Architecture

Describe the high-level architecture that requires reading multiple files to understand. Focus on:

• System boundaries (frontend/backend, microservices, monorepo packages)
• Request/data flow from entry point to storage
• Key abstractions, patterns, and conventions that are not obvious from a single file
• How modules/packages/services connect to each other
• Technology choices and why they matter for navigating the code

Keep it concise. This is the "big picture" overview, not a file listing.

Section 3: Conventions (conditional)

Header:

## Conventions

Only include this section if there are non-obvious project-specific conventions found in rules files, README, or CONTRIBUTING.md. Examples:

• Naming conventions that differ from language defaults
• Required patterns for new features (e.g., "every API endpoint needs a corresponding schema")
• Architectural constraints (e.g., "no direct database access from route handlers")
• Import ordering or module organization rules

If no meaningful conventions are found, omit this section entirely.

Anti-patterns -- DO NOT include any of the following

• Generic development advice ("write tests", "use meaningful names", "handle errors gracefully")
• Obvious instructions ("do not commit secrets", "provide helpful error messages")
• Exhaustive file trees or component listings (these are easily discoverable)
• Made-up sections like "Common Development Tasks", "Tips for Development", "Support and Documentation" unless content for them was explicitly found in project files
• Information that repeats what is already stated in another section
• Placeholder or aspirational content -- only document what actually exists

Phase 3: Create or Update AGENTS.md

If AGENTS.md does NOT exist

Create

AGENTS.md

at the repo root. This file is specifically for AI coding agents. Include:

• A brief one-liner about the project
• Agent-specific workflow hints (e.g., "run tests with X before committing", "use Y for database migrations")
• References to INIT.md for full context
• Any tool-specific notes extracted from existing rules files

Keep it short (under 50 lines). Agents will also read INIT.md for full context.

If AGENTS.md already exists

Read the existing file, then merge new findings:

• Add any missing commands or architecture notes as new sections
• Do not remove existing content
• Do not duplicate information already present
• Add a comment noting what was updated and when

Phase 4: Handle Existing CLAUDE.md

If CLAUDE.md exists

• Read it fully
• Compare its content against the generated INIT.md
• Suggest concrete improvements (missing commands, outdated architecture notes, redundant content)
• Present suggestions to the user -- do NOT overwrite CLAUDE.md

If CLAUDE.md does NOT exist

• No action needed. INIT.md serves the same purpose.

Quality Checklist

Before finishing, verify:

• INIT.md starts with the required preamble text
• All listed commands were extracted from actual project files (none invented)
• Architecture section describes the big picture, not individual files
• No generic/obvious development advice is included
• No made-up sections exist
• No repetition between sections
• AGENTS.md was created or updated
• If CLAUDE.md existed, improvement suggestions were provided