Agent workspace

Agent Workspace Starter Kit

v1.0 MIT AGENTS.md compatible

If you use AI coding agents you have said it this week: "Does anyone else spend twenty minutes every session re-explaining their project?" The agent is not broken — the repo is not legible to it. This kit fixes that with seven plain markdown files and one install script, so Claude Code, Cursor, Copilot, Codex, and Gemini all read your project state the same way.

License
MIT starter
Format
Plain markdown + scripts
Install
10 min one command
Price
Free Pro Pack soon

What it is

A file grammar your agent already understands.

AGENTS.md, CLAUDE.md, .cursorrules, and the rest each describe the agent's behavior — not the project's state. There is no shared grammar for what the project is, what's been decided, what's next. The kit adds exactly that layer.

01

One canonical spec

A single AGENTS.md is the source of truth. CLAUDE.md, GEMINI.md, .cursorrules, and the Copilot file are generated from it — so they cannot drift apart.

02

The project-state layer

PROJECT-STATE.md, DECISIONS.md, and CHANGELOG.md hold what every agent re-asks for: where the work is, what was decided, what changed.

03

Working scripts, not just templates

init-workspace installs the kit; kit-sync regenerates the derived files from the canonical spec; kit-status catches stale state, drift, and broken imports.

How it works

One workspace, every agent.

Each tool reads a different file natively — so the kit ships parallel files that all converge on the same project-state docs. Edit the canonical spec, run sync, and every tool stays in lockstep.

Drop it in

Three install paths

The GitHub "Use this template" button for new repos, a one-line install for existing projects, or a manual copy. Ten minutes either way.

Five tools

Claude, Cursor, Copilot, Codex, Gemini

Each reads its own native file; all of them follow the same breadcrumb trail to PROJECT-STATE.md and DECISIONS.md.

No lock-in

Plain markdown, plain folders

No hosted workspace, no database, no vendor account. The whole kit is files in your repo — version them, fork them, delete them.

Who it's for

If this sounds like you, it will fit.

  • A solo developer or indie hacker running AI coding agents on real projects, tired of "Groundhog Day" at the start of every session.
  • Anyone managing 1–5 active repos whose AGENTS.md, CLAUDE.md, and Cursor rules have already drifted apart.
  • People who switch between Claude Code, Cursor, Copilot, Codex, and Gemini and want one source of truth, not five files to hand-sync.

Plainly

What it does, and what it doesn't.

What it does

  • Gives five agent tools one shared, legible workspace structure
  • Keeps a canonical AGENTS.md synced to every tool's native file
  • Holds project state in PROJECT-STATE.md / DECISIONS.md / CHANGELOG.md
  • Ships working install, sync, and health-check scripts
  • Includes worked examples — a populated sample project and a before/after

What it doesn't

  • It is not for you if you use one tool and one rules file is enough
  • It is not a hosted Notion- or Obsidian-style workspace — plain markdown is the point
  • It is not SaaS drift-detection across a large engineering org
  • It will not write your project docs for you — it gives them a place and a grammar

Get it

Two ways in.

Start free with the starter kit — it stands on its own. The Pro Pack, in development, adds the kits and tooling for running many projects at once.

Free starter

Starter Kit

$0 · MIT

The seven-file workspace structure, genuinely useful on its own.

  • Canonical AGENTS.md + the four derived tool files
  • The PROJECT-STATE / DECISIONS / CHANGELOG layer
  • init / kit-sync / kit-status scripts
  • Starter templates + worked examples
  • A Claude Code skill that makes the agent kit-aware
Use the template on GitHub

"Use this template" — no account, no clock.

Stop re-explaining your project.

Drop the kit into your repo, fill in PROJECT-STATE.md once, and your next agent session opens with what you were doing — not "what is this project?"