Quick start

The stack ships as a Claude Code plugin. One install gives you four things: a project template, a set of agents that build and review work in a loop, a design loop, and a status line. This page installs the plugin and gives the first prompt for each; every one links on to the page that covers it in full.

Install the plugin

You need Claude Code. From inside it, run:

/plugin marketplace add MaybloomTech/maybloom-stack
/plugin install maybloom-stack@maybloom-stack

That installs every skill and every agent, namespaced under maybloom-stack (the template skill is maybloom-stack:maybloom-stack-bootstrap, the builder agent is maybloom-stack:builder). Skills says what each skill does, and how to update the plugin later.

You use skills by describing what you want in a session: a skill’s description lists the phrases it answers to. You use agents by launching them, as the pages below show.

What you need

Every use needs Claude Code and the plugin above. The rest depends on which use you want; each block is a checklist, and the page it links to has the detail.

The template. The scaffold builds and runs on your machine, so it needs:

• Python 3, which runs the skill’s scaffold script, and git, for the git init -b main the checklist ends with.
• Node 24 (the scaffold’s .nvmrc) and pnpm 10 (its packageManager pin, pnpm@10.26.2). pnpm install brings buf with it as an npm package; there is nothing else to install for the contract.
• Docker, for the Postgres the Go backend develops against (docker compose up -d db) and for scripts/go, which runs the pinned Go toolchain in a container. With Go 1.26.9 installed you can run go directly instead; with the TypeScript backend there is no database to start, because it uses PGlite.
• Decisions the skill asks you for, among them the display name, the slug, the npm scope, where to put the project, public or private, and the licence with its copyright holder. It scaffolds Go unless you name TypeScript (Choosing a runtime).

The skill’s checklist, and Skills, have the rest.

Agents and the review loop. These work through your project’s pull requests and its tracker, so decide these first:

• A forge and its command-line client. The agents are written for GitHub with gh (signed in, gh auth login) and for Forgejo with the forgejo skill’s fjo, which installs as its own plugin (how).
• A tracker, whose skill the agents load to read and comment on the item; the branch is named for the item’s id (PROJ-123/short-slug). The agents assume one; their files do not describe a run without it.
• A ## Delegation section in the project’s AGENTS.md: the protected paths, where commands run from and the checks a builder runs. The scaffold writes it; a project that did not start from the template writes its own.
• A permission mode for each launch: acceptEdits for the orchestrator and the builder, default for the reviewer.
• The status line below, because the budget hook reads its numbers.

Running a loop has the launch commands, the forge and tracker notes and the budget.

The design loop. A laptop with Chrome, and a scaffolded interface with its design board. Design loop lists what each role needs.

The status line. jq, curl and git, a signed-in claude.ai session for the usage segments, and Claude Code 2.1.251 or later for the cache countdown. Status line has the full list.

A new project from the template

Use this when you are starting a service and want the stack’s shape from the first commit: a protobuf contract, a Go backend (TypeScript on request), an Expo client and a docs site, with one example resource wired through every layer. In a session started in an empty folder, type:

Bootstrap a new maybloom stack project called Orchard.

The skill asks for the slug, the npm scope, where to put it, the backend, public or private, and the licence, then scaffolds the project and walks you through the checklist that follows. Go is the default; say with a TypeScript backend to get the other one. Once the project exists, ask for add a Book resource or add a due date to Note and the sibling skills take over. See Skills for the family and Choosing a runtime for the backend question.

Agents that build and review in a loop

Use this when the work is bigger than one conversation: an architect session writes a brief, a builder builds it, a fresh reviewer reviews it round by round. The first thing to type, in a session in your project, is:

Help me write a brief for this slice: <what you want built>.

The delegation skill answers write a brief with its brief template and sizes the slice to a session. To start a whole loop on a card, say:

Write the brief for PROJ-123 and delegate this slice: dispatch a builder
for it in its own worktree, review its pull request, and come back to me
only for a decision.

Running a loop has this prompt in full, and two more: one that hands the run to an orchestrator, one that starts with the design loop. To launch the builder yourself, the brief is its whole prompt:

claude --agent maybloom-stack:builder --permission-mode acceptEdits

Running a loop has the launch command for every role, one prompt to paste for each, a filled brief and a walk through one run from card to merged pull request. Delegation says why the work is split this way.

A design before it is built

Use this when a screen’s look needs deciding before anyone builds it: a designer proposes sketches and builds one pass on the design board, a critic judges it from screenshots, and you look last. It runs on a laptop with Chrome. In a session in a scaffolded project, type:

Run the design loop for the notes list screen.

The design loop page has the launch commands for the designer and the critic, the capture script and what happens after you approve.

The status line

Use this to see your usage and your session budget while you work: the 5-hour and weekly windows with their reset countdowns, context against the budget, and the prompt cache countdown. It also feeds the budget hook the scaffold installs. It needs jq, curl and git. In any session, type:

Install the status line.

The skill copies one script into ~/.claude and sets the statusLine key in ~/.claude/settings.json. The status line page shows the output, what the install changes, how to check it and what to do when a segment is missing.