Superpowers: a skill set that makes Claude Code plan and test before coding
TL;DR
Superpowers (obra/superpowers) is a library of workflows for Claude Code. Its skills load automatically and push the agent through steps good engineers already follow: clarify requirements, write a plan, code with TDD, debug systematically and verify before declaring completion.
The most common way to use AI for coding is to type "build me app X" and hope. That's fine for small things; for multi-file features it usually leads to code that runs but heads the wrong way, lacks tests and is hard to review. Superpowers (obra/superpowers) is a workflow library for Claude Code: skills load automatically and push the agent through the steps good developers already follow: clarify requirements, plan, code with TDD, debug systematically and verify before declaring done.
The main skills
| Skill | What it does |
|---|---|
| Brainstorming | Asks questions to clarify the idea before writing any code |
| Writing plans | Writes a detailed plan: phases, verification commands, success criteria, rollback options |
| Executing plans | Executes in batches, with review checkpoints |
| Test-driven development | RED → GREEN → REFACTOR: failing test first, code after |
| Systematic debugging | Find the root cause → find the pattern → test hypotheses → fix |
| Verification before completion | Must run verification commands and see the results before saying "done" |
Why this workflow works
- Early mistakes are cheap: fixing a paragraph of the plan is far cheaper than rewriting thirty files.
- Tests first clarify requirements: when the agent must write a failing test first, "what correct means" is defined by something runnable.
- Review checkpoints: batched execution gives humans a place to step in before mistakes spread.
- No self-declared completion: the verification step makes the agent prove it with command output.
Installation
/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplaceA typical workflow
- You describe the feature at the idea level.
- The agent asks clarifying questions (brainstorming) and summarises the design for you to confirm.
- The agent writes a plan file: phases, verification commands, success criteria, rollback.
- You review the plan: the most important control point.
- The agent executes in TDD batches, pausing for you to look after each one.
- Before saying it's done, the agent runs verification and shows the results.
When not to use the full process
- A one-line fix, a variable rename, a wording tweak: the full process costs more time and tokens than it's worth.
- Exploration and throwaway prototypes: a detailed plan is overkill.
- When you don't yet know what you want: use brainstorming to clarify, don't rush into a plan.
How to apply it
- Dev: use it for multi-file features, migrations and refactors, where missing a plan is expensive.
- Tech Lead: treat the plan file as a small design document; review it like a design doc before the agent touches code.
- PM: the plan and success criteria the agent writes can serve as acceptance criteria to check against at sign-off.