A universal multi-agent project harness for Claude Code.
Run any project — web, data, ML, fintech, CLI — through a structured AI team with rigorous governance, from discovery to release.
Table of Contents
What is atelier?
atelier is a Claude Code plugin that turns project execution into a structured multi-agent workflow. Instead of asking one Claude to "build me an app," you get:
- Five Phase 1 specialists who interview you, design the system, form a project-specific AI team, build the roadmap, and consolidate documentation — and then wait for your explicit approval before any code is written.
- Four Phase 2 specialists (a Software Engineer plus three independent reviewers) who implement tasks one-at-a-time through a PR / MR cycle with mandatory unanimous approval.
- Twelve hooks that hard-block the most common mistakes (force-push, skipped requirements, unreviewed merges) at the harness layer — the model cannot bypass them.
It works for any project domain (web app, data pipeline, fintech backend, ML system, CLI tool, mobile app) and any forge (GitHub, GitLab, Bitbucket, Gerrit, local-only).
Why atelier?
The problem
LLMs writing code are powerful, but in real projects they:
- Skip due diligence — start coding before requirements are clear.
- Lose context — forget design decisions across sessions or tasks.
- Drift in scope — silently expand a small task into a refactor.
- Approve their own work — single-agent setups have no independent review.
- Lose trace — decisions vanish into transcripts that no one re-reads.
The atelier approach
| Problem | atelier mitigation |
|---|
| Skipping due diligence | 6-file STEP gate: feature branches blocked until requirements + design + roadmap + approval all exist. |
| Lost context | Single Source of Truth in docs/ consulted by every agent. Glossary, ADRs, lessons-learned all append-only. |
| Scope drift | QA Engineer independently reviews every PR against original acceptance criteria. |
| Self-approval | Three mandatory reviewers (Senior Software Engineer, Tech Lead, QA Engineer) with independent lenses; unanimous required. |
| Lost decisions | ADR triggers mandate a written record for every consequential choice. Tech Lead blocks merges that fire a trigger without an ADR. |
vs alternatives
| Setup | What you get | What you give up |
|---|
| Single-agent prompting | Speed, simplicity | Independent review, governance, trace |
| Manual multi-agent (custom prompts each time) | Flexibility | Reproducibility, cross-project consistency, gates |
| atelier | Structured team + governance + gates + trace | Some upfront ceremony at STEP 0.5 (offset by reusability) |
Quickstart
Prerequisites
- Claude Code installed and authenticated.
- A directory where you want to start a new project (or retrofit an existing one).
- Optionally:
gh CLI (GitHub) or glab CLI (GitLab) for forge integration. atelier-open-pr will detect which one to use.
Install
From the marketplace (recommended) — atelier is self-hosted as a single-plugin marketplace from this repo:
/plugin marketplace add dudgns0908/atelier
/plugin install atelier@atelier
The form is <plugin-name>@<marketplace-name>; both happen to be atelier.
