Claude Code Superpowers: setup and how it works

Superpowers is the most widely used skill collection for Claude Code, with close to 300,000 GitHub stars in October 2026. It is not one skill but a working method: before Claude writes code it asks what you are trying to do, writes a short design, breaks the work into small tasks, builds them test-first and reviews the result. The skills trigger on their own, so after installing it you mostly notice that Claude stopped jumping straight into code.

This guide covers installation, the workflow step by step, what each skill does, what it costs and when to leave it off. If skills are new to you, start with what Claude skills are; for other collections worth installing, see the best Claude skills.

Facts below come from the obra/superpowers repository, checked on 12 October 2026.

Install Superpowers in Claude Code

Superpowers is in Anthropic's official plugin marketplace:

/plugin install superpowers@claude-plugins-official

The author's own marketplace carries it too, along with related plugins:

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

To check it is active, start describing something you want built: instead of writing code, Claude should start asking questions. That is the brainstorming skill.

Superpowers also installs in Codex, Cursor (/add-plugin superpowers), Gemini CLI, GitHub Copilot CLI, OpenCode, Devin CLI and others. Each tool needs its own install.

The workflow, step by step

The core of Superpowers is seven skills that hand over to each other:

1. brainstorming starts before anything is built. Claude asks what you want and why, then plays its understanding back so you can correct it. Small changes are agreed in chat; larger ones get a written design document.
2. using-git-worktrees starts once the design is approved. It creates an isolated workspace on a new branch, runs the project setup and checks that the tests pass before anything changes.
3. writing-plans turns the design into tasks of two to five minutes each, with exact file paths, the code to write and how to verify it.
4. subagent-driven-development or executing-plans carries out the plan. The first gives each task to a fresh subagent and reviews after every task (most thorough). The second runs the whole plan in the current session with one review at the end (cheaper).
5. test-driven-development applies during the work: write a failing test, watch it fail, write the minimum code, watch it pass, commit. Code written before its test gets deleted.
6. requesting-code-review runs between tasks, checks the work against the plan and blocks progress on critical issues.
7. finishing-a-development-branch verifies the tests and offers to merge, open a pull request, keep the branch or throw it away, then cleans up the worktree.

The repository describes these as "mandatory workflows, not suggestions": the agent checks for a relevant skill before every task. That is the point of it, and also the main reason some people turn it off.

Every skill in the collection

What it costs

Superpowers trades tokens for fewer wrong turns. A design, a plan, failing tests first and a review per task all cost tokens that "just fix it" does not.

• Subagent-driven development is the expensive mode. Each task gets a fresh subagent that has to read its way into the code, plus a review. In return, the main session's context stays clean, which is what lets it run for hours.
• executing-plans is the cheaper mode. One session, one review at the end. Good for plans of a handful of tasks.
• On a Pro plan you will hit limits sooner. If you pay per token, our Claude API pricing guide shows what a long agent session costs; on subscriptions, see Claude Code pricing.

The author's README says it is "not uncommon" for the agent to work autonomously for a couple of hours from an approved plan. Budget for that before you let it.

When Superpowers is the wrong tool

• One-line fixes and quick questions. The process is overhead for a typo. Say it is a small change, or disable the plugin for that session.
• Exploratory spikes. Test-first work slows you down when you do not yet know what you are building. Brainstorm, then switch it off while you prototype.
• Another methodology already installed. Two collections with different rules for when to plan and when to test give Claude conflicting instructions. Pick one.
• Repositories without a test setup. test-driven-development assumes you can run tests. Set that up first, or the workflow stalls at step 2.

When a session goes wrong (a skill fires when it should not, or the agent ignores its plan), ask Claude to "figure out what went wrong with superpowers in this session". The diagnosing-superpowers skill reads the transcript and points to the cause.

Superpowers across a team

Superpowers pays off most when everyone uses it. Then every developer's agent plans, tests and reviews the same way, and pull requests from different people look alike.

That raises the usual team questions: which version is approved, who updates it, and whether the agent may merge on its own. Pin the plugin version, review updates before rolling them out, and decide whether finishing-a-development-branch may merge or only open pull requests. Our skills guide covers rolling skills out to a team in more detail, and Walma AI Hub distributes one approved set of skills and MCP servers to every developer's Claude client, with one policy and one log, hosted in the EU. See how it works.