Superpowers for Claude Code explained
AI coding agents are fast. The trouble is the order they work in: they start typing before they know what you want, skip the tests and announce "done" without checking.
Superpowers goes after that habit. It doesn't add a team of robot employees. It adds a process: ask, plan, test first, review, then prove it works.
01What Superpowers actually is
Superpowers is an open-source project by Jesse Vincent and the team at Prime Radiant, released under the MIT license. Its README calls it "a complete software development methodology for your coding agents, built on top of a set of composable skills."
A skill is a plain markdown file (named SKILL.md) with a name, a "use when" description and step-by-step instructions. When a situation matches the description, the agent loads the file and follows it. Skills are mostly instructions, not software, though a few bundle small helper scripts.
Vincent announced it on October 9, 2025, the day Anthropic rolled out its plugin system for Claude Code. Version 7.0.0 shipped on October 9, 2026, and the GitHub repo showed roughly 300,000 stars when we checked on October 20, 2026. It is no longer Claude Code only: the README has install steps for 16 setups, including Codex, Cursor, Gemini CLI, GitHub Copilot CLI and Factory Droid.
02The 15 skills, by category
The README sorts the skills into four groups. Most are about collaboration: how the agent works with you and its helpers.
Chart 1
What the 15 skills cover
| Category | Skills | Skill names |
|---|---|---|
| Collaboration working with you and helpers | brainstorming, writing-plans, executing-plans, subagent-driven-development, dispatching-parallel-agents, requesting-code-review, receiving-code-review, using-git-worktrees, finishing-a-development-branch | |
| Debugging finding and proving fixes | systematic-debugging, verification-before-completion, diagnosing-superpowers | |
| Meta skills about skills | using-superpowers, writing-skills | |
| Testing test first, always | test-driven-development |
skills/ at v7.0.0 (October 9, 2026).Three skills carry what the project calls an "Iron Law", written in capitals so the agent can't miss it:
- test-driven-development: "NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST." Test-driven development (TDD) means you write a small automated check first, watch it fail, then write just enough code to make it pass.
- systematic-debugging: "NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST." It walks four phases: investigate the root cause, compare with working patterns, test one hypothesis, then fix.
- verification-before-completion: "NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE." The agent has to run the checks and read the output before it says something works.
03How it switches on: no commands to learn
Most plugins give you slash commands to remember. Superpowers works the other way round. When a session starts, is cleared or is compacted (shortened to save memory), a small startup script loads the using-superpowers skill into the conversation. That skill tells the agent to check for a relevant skill before it responds to anything.
So you just talk normally. Say "let's build a signup page" and the brainstorming skill should kick in before any code is written. The project even removed its last three legacy slash commands in v5.1.0 (April 2026), because they only pointed at the skills anyway.
The wording is deliberate. In his launch post, Vincent wrote that the skills borrow persuasion principles from Robert Cialdini's Influence, and that he tests them on helper agents under realistic pressure, like a fake outage.
04The core workflow, step by step
The README lays out a "basic workflow" that chains the skills together. Each step activates on its own when the previous one finishes.
Chart 2
From fuzzy idea to merged branch
brainstormingDraws out what you want and why, then plays it back for correction.using-git-worktreesSets up a separate workspace on a new branch and checks the tests pass.writing-plansBreaks the approved design into small tasks with files and checks.subagent-driven-developmentA fresh helper builds each task test-first, then a review follows.finishing-a-development-branchRuns the tests and offers merge, pull request, keep or discard.
Cheaper option at step 4: executing-plans builds every task in the main session and runs one review at the end.
The jargon, in plain English
A few terms worth unpacking. A git worktree is a second working folder for the same project, on its own branch, so experiments can't break your main code. A subagent is a fresh helper session with a clean memory, handed one task. Each task gets its own builder, then a review for "does it match the spec?" and "is the code any good?". Fix loops are capped at five rounds.
The plan is the heart of it. The README says plans should be clear enough for "an enthusiastic junior engineer with poor taste, no judgement, no project context, and an aversion to testing" to follow. It's a useful test for any spec you write yourself.
05What's proven vs. what's claimed
Superpowers' release notes report results from the project's own evaluations. Useful, but not independent benchmarks.
Chart 3
How well each claim is backed up
| Claim | What we found | Evidence |
|---|---|---|
| Skills trigger automatically | A startup hook loads using-superpowers on start, clear and compact. | Verified in code |
| 15 skills | Counted the folders in skills/ at v7.0.0. | Verified in code |
| Skill changes are tested | Release notes publish the project's own eval results, such as leaner plans in v6.4.2. | Project's own tests |
| Zero dependencies | Stated as a design rule in the contributor guide. | Documented |
| Hours of autonomous work | The README says this is "not uncommon". No published measurement. | Claim only |
One example: the v6.4.2 notes say rewritten planning made plans take "a quarter of the time and about a third of the tokens" on a reproduced bug report. Specific, but measured by the maintainers on their own scenario.
06Superpowers vs. oh-my-claudecode
We covered oh-my-claudecode (OMC) separately, and the two get compared a lot. OMC even lists Superpowers among its inspirations. The difference is philosophy: Superpowers changes how one agent works; OMC organises many agents into a team.
Chart 4
Habits vs. a team
| Feature | Superpowers | oh-my-claudecode |
|---|---|---|
| Core idea | Skills that change how one agent works | 19 specialist agents plus workflow modes |
| Skills | 15 | ~48 |
| Commands to learn | NoneSkills trigger from plain requests | A few/autopilot, /team, /deep-interview |
| Named specialist agents | NoHelper prompts for builder and reviewer | Yes19, each pinned to a model tier |
| Parallel agents | PartlyOne skill for independent tasks | YesTeam pipeline with verify and fix |
| Test-first rule | YesAn "Iron Law" in the TDD skill | Opt-inA tdd skill plus test-engineer and verifier agents |
| Extra tools needed | None by design (Git Bash on Windows) | tmux for some features |
| Recent breaking change | v5.1.0 removed 3 legacy commands | v5.0 removed 14 skills and 7 commands |
| License | MIT | MIT |
| GitHub stars, Oct 2026 | ~300k (repo page) | ~40k (project site) |
Pick Superpowers for discipline with almost nothing to learn. Pick OMC to throw several agents at a big, clear job. Both hook into your sessions, so try one at a time.
07Why this matters if you're a founder in the making
Solo founders rarely have someone asking "what problem are we solving?" before work starts. Superpowers builds that question in.
- It interviews you first. The brainstorming skill's own description says to use it before any creative work, "in software or out of it (a talk, a business, a renovation)". For non-specialists, that's the most valuable part.
- Tests become a safety net. If you can't read every line, passing tests tell you yesterday's feature still works.
- "Done" means proven. The verification skill makes the agent show evidence, not just say "fixed".
- You can teach it your way. The
writing-skillsskill helps you turn your own processes into new skills.
If you're brand new to Claude Code, start with our Claude Code beginner's guide, then add Superpowers once you're comfortable.
08The catches
- It's strict on purpose. The README says the TDD skill "deletes code written before tests". Great for a product, annoying for a throwaway prototype. Version 7.0.0 says brainstorming now sizes the work to the job, but expect questions.
- It spends tokens on process. A fresh helper and a review per task cost more than one agent typing. The README calls
executing-plansthe cheapest option. - It changes often. The release notes list 45 versions between October 2025 and October 2026, and v7.0.0 rebuilt brainstorming from scratch. Older tutorials may show commands that no longer exist.
- Optional telemetry. The brainstorming "visual companion" loads a logo from Prime Radiant's website that includes your Superpowers version. Set
SUPERPOWERS_DISABLE_TELEMETRYto turn it off. - Core is closed to new skills. The maintainers say they don't generally accept new skill contributions. Custom skills belong in your own plugin.
09How to try it on a side project
You need Claude Code installed and signed in. Inside a Claude Code session, install from Anthropic's official plugin marketplace:
/plugin install superpowers@claude-plugins-officialOr use Jesse Vincent's own marketplace. Enter these one at a time:
/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplaceStart a fresh session and describe something real. A working install should start brainstorming before any code appears:
Let's build a waitlist page for my app with email signupIf a session goes sideways, ask: "figure out what went wrong with superpowers in this session". That triggers the diagnosing-superpowers skill, which reads the session log and reports what happened.
10Verdict
Superpowers is the lightweight way to make an AI coding agent behave like a careful engineer. It adds no dashboards and no agent roster, just 15 playbooks, tested by the maintainers' own evals, that switch on when needed. It's free and MIT-licensed.
The cost is patience: more questions, more tests, more tokens per feature. For a product customers will touch, that trade is usually worth it. Try it on a side project and judge it on whether the result actually works.
SourcesSources and further reading
- Superpowers on GitHub: README,
skills/folder, hooks, AGENTS.md and LICENSE (checked at v7.0.0, commit bb92a77, October 2026; star count from the repo page on October 20, 2026). - Superpowers release notes: v7.0.0 (October 9, 2026), v6.4.2 eval figures, and v5.1.0 removal of legacy commands.
- Superpowers: How I'm using coding agents in October 2025: Jesse Vincent's launch post.
- Superpowers in the official Claude plugin marketplace: as linked from the README.
- oh-my-claudecode on GitHub: comparison facts (v5.6.1).
- Claude Code documentation: installation and plugins.
Facts were checked on October 20, 2026. Tools like this change quickly; if something here is out of date, tell us on X.
Your Average Tech Founder
Host of The relatable entrepreneurship podcast, recorded in Da Nang, Vietnam. Writes about the tools, repos and habits that help founders in the making build better. No guru energy.