1. Install the plugin
Every developer installs Huddle once, on their own machine. A settings file committed to the repo does not do it for them. You need:
- Claude Code (opens in a new tab).
- Node 22 or newer, and
git. - On GitHub, the gh (opens in a new tab) command line, logged in with
gh auth login. On GitLab, the glab (opens in a new tab) command line, logged in withglab auth login(add--hostname gitlab.example.comfor a self-managed server).
Run this once in any terminal, then open a new Claude Code session:
claude plugin marketplace add SrProfessional/Huddle
claude plugin install huddle@huddle
In the Claude Code command line you can type the same as /plugin marketplace add SrProfessional/Huddle and /plugin install huddle@huddle. The VS Code extension uses /plugins instead of /plugin: type /plugins, add the marketplace SrProfessional/Huddle in the Marketplaces tab, install huddle from the Plugins tab with Install for this project (not «Install for you»), and type /reload-plugins (or restart Claude Code). Each person does this once per clone. Huddle has no runtime dependencies, so there is nothing to npm install.
To update later. Install and update are per project, so repeat this in each project that uses Huddle. From the project, type one command:
/huddle-update
It updates the plugin (claude plugin marketplace update huddle and claude plugin update huddle@huddle --scope project), refreshes the templates and the CLAUDE.md block in a branch of its own, asks only the questions a new release added, and opens the PR. It ends by telling you to type /reload-plugins; if the new skills do not show up, restart Claude Code. A project can be on an older version than another. /huddle-play and /huddle-update tell you in the chat when a newer release exists, at most once a day.
2. Set up your repo
This is the AI Manager's step, once per repo. Open Claude Code in the repo and create a branch for the setup:
git switch -c huddle-setup
- Run
/huddle-setup. It asks for the platform (GitHub or GitLab), the host (empty for github.com or gitlab.com), the project, the login of the AI Manager and the logins of the developers. - It writes
huddle.json, the Issue and pull request templates, the job that keeps the labels up to date, and the Huddle block ofCLAUDE.md. Nothing outside that block is touched. - Run
huddle doctor. You want a check mark on the file, the login, the AI Manager and the labels. If a label is missing, runhuddle labelsonce. - Commit, push and open a pull request with the setup files. Merge it. The AI Manager is the only one who merges.
On GitLab, two more steps are only the AI Manager's: a pipeline schedule with the cron */10 * * * * on the default branch, and a project access token with the api scope stored as the masked, protected CI variable HUDDLE_TOKEN. Labels can take up to ten minutes to change. The GitLab notes (opens in a new tab) explain both in detail.
3. Write the first Issue
Run /huddle-issue. It writes the Issue from the template, so every task arrives complete:
| Part | What it says |
|---|---|
| Goal | What we want to achieve and why, in a few lines. |
| Prompt for Claude Code | The text the developer pastes, with the steps, the branch and how to open the pull request. |
| Allowed files | The only files the session may touch. |
| Forbidden files | What is not touched. If it must be, the session asks first. |
| Contract or spec | The document that rules over the code. |
| Tests | What must pass before the pull request opens. |
4. A developer's first day
- Set up your machine. Install Claude Code, Node,
gitandghorglab, log in yourself, and install Huddle with the two commands above. Never paste a token into a chat, a file or an Issue. - Check it works. Clone the team's repo, open Claude Code in it and run
huddle doctor. Iflabelsfails, tell the AI Manager: only they runhuddle labels. - Take an Issue. The AI Manager assigns one, or the labels job does. Run
/huddle-play <n>with its number, or paste the Issue's prompt. The session checks what the Issue depends on, creates its own branch and adds thein progresslabel. If a dependency is not merged yet, it stops, and that is right. - Answer its questions. This is the job. If you do not know, say so and let the session ask the AI Manager.
- Review what it did before you accept the pull request. You are responsible for it.
A session ends with a message that starts with «Issue #N finished» or «Issue #N stopped», with the pull request link, the test counts, what was not done and your next step. After the pull request opens, the session waits in the background and picks up the AI Manager's requested changes on its own.
Next: how it works, with the roles, the labels and the reviews.