AGENTS.md Generator: Context Files for Coding Agents

Repository AGENTS.md Generator is an agent skill that writes AGENTS.md files for the folders in a code repository — the file coding agents read to learn how this project builds, tests, and behaves. It comes from Microsoft's public skills repository, and its first rule is what makes it safe to run on a real codebase: it never overwrites an AGENTS.md that already exists.

What AGENTS.md is

AGENTS.md sits alongside your README, with a different audience. The README is for humans; AGENTS.md is for coding agents. Agents look for it in the folder they're working in and walk up the tree, and a subfolder can have its own file that takes precedence over the root one. That gives exact commands, conventions, and limits a home that doesn't clutter the README.

The skill also heads off a common mix-up. AGENTS.md is not an agent persona file. A persona describes who the agent is; AGENTS.md describes what the agent should know about this code.

What it does

It checks before it writes. For every folder, independently, it looks for an existing AGENTS.md. If one is there, it skips that folder and reports the skip. Nothing you've tuned by hand gets replaced.

It picks the folders that deserve one. Always the repository root. Then src/, tests/, app/, or api/ if they exist, each package in a monorepo, and any folder with its own build manifest — package.json, pyproject.toml, Cargo.toml, go.mod, and the like. It always skips dependency and build-output folders such as node_modules/, dist/, and virtual environments.

It reads the real project. Commands come from package.json scripts, Makefile targets, pyproject.toml, and CI workflow files. Conventions come from reading three to five actual source files for naming, import order, and error handling.

It covers six areas, only where they apply. Build and run commands go first, because agents reach for them constantly. Then testing, project structure, code style with one real example, git workflow, and boundaries. A folder with no tests gets no testing section; the skill is explicit about not inventing sections for things a project doesn't have.

It validates before saving. Every command must point at a real script or tool, every path at a real file, and no placeholder text like TODO is allowed to survive.

The boundaries section

This is the part teams get the most out of. The skill writes limits in three tiers: always do (run tests before committing), ask first (add a dependency, change a database schema, touch CI), and never do (commit secrets, edit vendored code, push straight to main). It tailors them to the folder — schema and API-contract rules for a backend, component-API rules for a frontend, production-config rules for infrastructure.

Those are the instructions you'd otherwise repeat to an agent at the start of every session.

When to reach for it

• Coding agents keep guessing the wrong build or test command in your repository
• You're bringing agents onto an existing codebase that has no agent context files
• A monorepo where every package builds and tests differently
• You want your "don't touch this" rules written down once
• New contributors, human or agent, need the exact commands quickly

It generates; it doesn't audit. If your repository already has AGENTS.md files you're happy with, it will skip them — which is the correct behaviour, and also means it won't improve them.

What's actually inside

One skill. No connectors, no API keys. Upstream, it's one of several skills in Microsoft's deep-wiki plugin, which also generates wikis, changelogs, and onboarding guides from a repository. The Taku listing publishes this one on its own, and adds that it works only from the materials you supply and labels assumptions and missing evidence.

Setting it up

The GitHub route. The skill lives at .github/plugins/deep-wiki/skills/wiki-agents-md in microsoft/skills — inside a hidden .github folder, several levels down, inside a plugin. Copy that folder into your agent's skills directory (for Claude Code, .claude/skills/), or install the whole deep-wiki plugin, then start a fresh session inside the repository you want documented. The project's documentation site covers the rest of the collection.

The Taku route. Open it in Taku and the skill is mirrored into your workspace in one step.

You still need a repository and a coding agent that can read it. What Taku removes is finding and installing the skill, not the codebase.

Who made it

Written by Microsoft and published under the MIT license in its public skills repository, as part of the deep-wiki plugin. Published to the Taku marketplace by n1ko.

If you've never installed a skill before

You don't need to understand GitHub, Codex, or Claude Code to get this skill running. Those routes work, but this skill sits in a hidden folder inside a plugin, which is exactly the kind of path that stops people before they start.

Taku is a desktop workspace built for that gap: find something someone already built, mirror it into your own workspace, and run it without the setup work. Browse the free app library for more developer skills, or read coding agents for why project context matters so much to how they perform. To picture the same repository instead of describing it, Archify turns a system into a checked, interactive diagram.

Download Taku to give your repository agent context without the install step. Taku is in Beta, and the Mac app is available now; there's also an experimental, unsigned Windows build for Windows 10 and 11.

FAQ

What is AGENTS.md?

A Markdown file of project context for coding agents: build and test commands, structure, code style, and boundaries. It complements README.md, which is written for humans.

Will it overwrite my existing AGENTS.md?

No. It checks every folder first and skips any folder that already has one.

Which folders get an AGENTS.md?

The repository root always, plus common source and test folders, monorepo packages, and any folder with its own build manifest. Dependency and build-output folders are always skipped.

Does it make up commands?

It's designed not to. Commands come from your package scripts, Makefiles, and CI files, and it validates that every command and path is real before writing.

Is it free?

The skill is free and open source. You need your own coding agent and a repository. For what Taku costs, see pricing.

Key points

• Writes AGENTS.md context files for coding agents, folder by folder.
• Never overwrites an existing AGENTS.md — it checks every folder first.
• Pulls real commands from scripts and CI, and conventions from real source files.
• Writes three-tier boundaries: always do, ask first, never do.
• From Microsoft's deep-wiki plugin; Taku removes the install, not the repository.