Superpowers: the skill-driven development methodology for Claude Code

At a glance

The skills split into three groups: testing & debugging (test-driven-development, systematic-debugging, verification-before-completion), collaboration (brainstorming, writing-plans, executing-plans, subagent-driven-development, requesting/receiving-code-review, using-git-worktrees, finishing-a-development-branch, dispatching-parallel-agents), and a small meta layer (writing-skills, using-superpowers) that teaches the agent how to extend the system with its own new skills.

The workflow: brainstorm → plan → TDD → review

Nothing here is invoked by name — a short set of always-on instructions tells the agent to check for a matching skill before starting any task, so the right stage triggers based on what you actually asked for:

1. brainstorming — Refuses to write code until it has asked clarifying questions and proposed design options for you to accept
2. writing-plans — Breaks the accepted design into 2–5 minute tasks with exact file paths and a verification step for each
3. test-driven-development — Per task: write a failing test, watch it fail, write the minimal code to pass, watch it pass, commit
4. subagent-driven-development — Dispatches a fresh subagent per task with only that task's spec, then runs a two-stage review — spec compliance, then code quality

The systematic-debugging skill runs the same discipline in reverse for bugs: a four-phase root-cause process instead of guess-and-patch. And verification-before-completion is the guardrail that stops an agent from marking a task done on the strength of "the code looks right" — it has to actually run the check.

RED      write a test that fails for the right reason
         → run it, confirm it fails
GREEN    write the minimal code to make it pass
         → run it, confirm it passes
REFACTOR clean up now that the behavior is locked in
COMMIT   one task, one commit, verification step attached

Installation

Official Claude Code marketplace, one line:

/plugin install superpowers@claude-plugins-official

Or straight from obra's own marketplace, which tends to get updates first:

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

The repo also ships plugin variants for Cursor, GitHub Copilot CLI, and Gemini CLI (its own .cursor-plugin, .codex-plugin, and .opencode directories are committed alongside the Claude Code one) — the skills themselves are plain Markdown, so the methodology travels even where the install mechanics differ. License is MIT throughout.

Using it day to day

You don't drive Superpowers with commands — you drive it by asking for the work you want ("add rate limiting to the API", "fix the flaky checkout test") and the initial instructions make the agent check its skill list before touching code. For a genuinely new feature, expect a real back-and-forth on the design before you see a diff; that friction is the entire point.

It's overkill for a one-line fix — a typo correction doesn't need a brainstorming pass. Where it earns its keep is anything at "build me a feature" scale, especially with subagent-driven-development: each task in the plan gets a fresh subagent with only that task's spec, then two review passes (spec compliance, then code quality) before it's considered done — which keeps a single agent from slowly accumulating an entire feature's context and drifting.

My take

The part that actually changes behavior isn't any single skill — it's that the instructions make the agent check its own workflow before starting, so drift from "we agreed to write a test first" becomes visible immediately instead of three files later. That's a genuinely different failure mode than the usual "the agent went off and did its own thing."

The honest tradeoff: it slows down small changes, and a team already running its own strict TDD/PR discipline may find the brainstorming and planning stages redundant with process they already have. Try it on a feature-sized task before deciding whether to keep it always-on.

Frequently asked questions