Skip to content

Guide · Agent skill

How to use the Compose Next agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

Compose Next is an agent skill from the MiMo-Code project (MIT licence) that gives an AI coding agent a structured process for larger changes: settling requirements, isolating work in a git worktree, writing a feature design document with tasks, then implementing, verifying and reviewing. This guide explains what the skill does, how to install it, how to trigger it with plain requests, and what it needs before you rely on it. By the end you will be able to decide whether it fits a change and run it safely.

Compose Next WorkflowXiaomiMiMo/MiMo-Code · MIT

Quick start

  1. 1Download the ZIP. Download the compose-next ZIP from this page.
  2. 2Unzip into a skills folder. Unzip it into ~/.claude/skills/ for personal use or into the project's .claude/skills/ folder for one repository.
  3. 3Or upload in the Claude apps. Go to Settings, then Capabilities, then Skills, and upload the ZIP.
  4. 4Check git and dependency tooling. Make sure git with worktree support and the repository's dependency tooling, such as bun or uv, are available.
  5. 5Ask explicitly for the workflow. Type /compose-next or name the skill in a clear request describing the change.
  6. 6Answer the decision questions. Respond to the agent's one-at-a-time questions about requirements and scope.
  7. 7Review the spec and results. Check the feature document, tasks and final report before merging the branch.

What the Compose Next skill makes the agent do

The skill is for multi-step feature work, bug fixes or refactors where requirements need to settle, where the feature should be documented, and where the change deserves independent review before merge.

When explicitly requested, the agent asks questions one decision at a time, sets up an isolated git worktree on a new branch, writes a single feature design document with tasks and acceptance criteria, then implements, verifies and reviews the work. The skill's stated phases are grill, workspace, spec, implement, verify, review, finalize and finish.

  • Requirements are settled through one-question-at-a-time decisions.
  • Changes go into a linked git worktree on a new branch, by default at .worktrees/{slug}.
  • One feature document holds design, tasks and acceptance criteria, by default at docs/compose/spec/{feature-name}.md.
  • Without an explicit request, the agent does the work directly and skips this process.

Install the compose-next skill

Download the compose-next ZIP from this page. Then choose one of the install routes below.

For a personal install, unzip the skill into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

  • Personal: ~/.claude/skills/
  • Project only: .claude/skills/ inside the repository
  • Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Trigger the skill with a natural request

The skill runs only when you explicitly ask for this workflow. You can type /compose-next, name the skill, or use any other clear natural-language request for the workflow. You do not need to know the skill's internal name.

The skill does not start just because a task is large. A one-shot edit, a single-file tweak or a plain question will not trigger it, and the skill says those need no orchestration overhead.

  • /compose-next add CSV export to the reports page
  • Use the compose-next workflow to fix the login redirect bug
  • Run this refactor through the Compose workflow with a design doc
/compose-next

Example: a feature with a design document

After you ask for the workflow, the agent first inspects the repository, its instructions (such as AGENTS.md or README) and existing spec files. It then asks only the questions the environment cannot answer, one decision at a time, and does not start implementation until requirements and scope are settled.

When the design is ready, the spec file holds problem, design, out-of-scope items and a task list in which each task has an observable acceptance criterion. The Report section stays empty until the work is finalized.

  • Settled decisions and contracts are recorded; exploration history is not.
  • Tasks can declare depends: links, which must not form cycles.
  • Small changes get a small document; the skill says not to pad designs.

Requirements and safety notes

The skill needs git with worktree support. It also needs the project's own dependency tooling, for example bun or uv, and it prefers lockfile-frozen install commands such as bun ci or uv sync --frozen. A question tool or interactive input is preferred but not required.

The skill does not run bundled scripts and does not use the network, according to its listing. It does write files and run git commands inside your project, so avoid it in repositories where you do not want the agent creating worktrees, docs/ folders or branches.

The skill will not begin implementation on main or master without your explicit consent. If you say without worktree or name a workspace path, it uses that choice.

  • Requires: git with worktree support
  • Requires: project dependency tooling as defined by the repository
  • Runs scripts: no
  • Uses network: no

Tips for getting good results

Give the agent a clear goal and say whether you want a spec. If you say without spec or this is a small fix, the skill skips the feature document and its spec gate, but keeps verification and review where the task warrants them.

If a project has its own worktree path, branch convention or spec path in AGENTS.md or user configuration, the skill uses that instead of the defaults and records the override.

  • State the workflow explicitly so it is not triggered by accident.
  • Answer the single decision questions the agent asks; unrelated decisions come in separate turns.
  • Review the spec before implementation if you are available to approve it.

What people use it for

  • Adding a multi-step feature that needs a written design and task list before coding.
  • Fixing a bug or refactoring a module where requirements need to be settled first.
  • Keeping changes isolated in a git worktree on a new branch until review.
  • Producing a spec file that records design, tasks and acceptance criteria alongside the code.

Frequently asked questions

Is the compose-next skill free to use?

The skill is released under the MIT licence, as listed on this page. Check the MiMo-Code repository for the licence text before redistributing it.

What do I need to run compose-next?

You need git with worktree support and the repository's own dependency tooling, such as bun or uv. A question tool or interactive input is preferred but not strictly required.

How do I trigger the compose-next skill?

Type /compose-next or clearly ask for the workflow by name or in natural language. The skill is designed not to run just because a task looks complex.

Does compose-next run scripts or use the internet?

According to this listing, it runs no scripts and uses no network. It does write files and run git commands in your project.

When should I not use compose-next?

Skip it for quick one-shot edits, single-file tweaks and plain questions, which the skill says need no orchestration overhead. Also avoid it in repositories where you do not want worktrees, docs/ folders or branches created.

Where do I install the skill?

Unzip the downloaded ZIP into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder, or upload it under Settings, Capabilities, Skills in the Claude apps.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Compose Next suits developers who want a repeatable process with a written design, a task list and an isolated worktree for non-trivial changes. It is a poor fit for quick edits, simple questions, or repositories where the agent should not create branches, worktrees or documentation folders; in those cases, ask for the change directly without the workflow.

Download the skill