Kuala Lumpur
← Back to the library

REPOVibe coding

Superpowers: make your coding agent plan and test before it builds

A free skill pack for Claude Code, Codex and Cursor. Your agent asks what you want, writes a short design, plans small tested steps, then builds and reviews its own work.

WHAT YOU’LL GET

Features built from a plan you approved, in small steps with tests, instead of a big guess you have to untangle.

WHO IT’S FOR

Anyone building an app or feature with Claude Code, Codex or Cursor who has had the agent build the wrong thing or call it done too early.

DIFFICULTY

Intermediate

TIME

About 10 minutes to install

WORKS WITH

Claude Codex

Get the full file

The whole resource as one Markdown file for your notes or your AI workspace.

FREEOpen on GitHub

What it does

Ask a coding agent for a feature and it usually starts typing code straight away. Sometimes that works. Often it builds something close to what you meant, and you spend the afternoon explaining the difference.

Superpowers is a set of skills by Jesse Vincent and the team at Prime Radiant that changes the order of work. When the agent sees you are building something, it first asks questions until it understands the goal, shows you the design in short pieces you can actually read, and waits for your OK. Then it writes a build plan of small tasks, each with the files to touch and how to check it. Only then does it build, with tests written first and a review after each step.

You do not have to call the skills by name. They switch on by themselves when the work fits.

Who it is for: people building real features with an agent. Who it is not for: a one-line fix. For small changes the full routine is too much; just tell the agent to skip it.

I have it installed in my own Claude Code (version 6.4.1 on 30 Sep 2026).

What you need first

  • Claude Code, Codex, Cursor or another supported agent. The upstream README lists more than a dozen.
  • A project in git. Superpowers works on a separate branch so your main copy stays clean.
  • Some kind of test setup helps a lot, because the build step writes tests first.
  • Superpowers is free, but it runs inside a coding agent such as Claude Code or Codex, which may need a paid plan.

Install (checked 30 Sep 2026)

Claude Code. Inside a Claude Code session, type:

/plugin install superpowers@claude-plugins-official

The same install from a normal terminal is claude plugin install superpowers@claude-plugins-official. We ran that on a throwaway setup on 30 Sep 2026: it installed version 6.4.1 with 15 skills. On that fresh setup we first had to add Anthropic's plugin list with claude plugin marketplace add anthropics/claude-plugins-official; your Claude Code may already have it.

Codex app. Open Plugins in the sidebar, find Superpowers in the Coding section and click the plus.

Codex CLI. Type /plugins, search for superpowers, and choose Install Plugin.

Start a new session afterwards so the skills load.

First use

Open your project and describe a real feature in plain words. For example:

What to expect, going by the upstream README:

  1. Questions first. The brainstorming skill asks what you are really after: where the enquiries should go, what counts as a valid phone number, what happens on an error.
  2. A design in short sections. You approve or correct each part. It saves the design as a document.
  3. A plan of small tasks. Each task names the exact files and the check that proves it works.
  4. Build when you say go. Either one fresh helper agent per task with a review after each (most thorough), or everything in the current session with one review at the end (cheaper).
  5. Finish. It runs the tests and asks whether to merge, open a pull request, keep the branch or throw it away. That choice stays with you.

Three skills worth knowing by name, because you can ask for them directly: systematic-debugging when something is broken and you do not know why, verification-before-completion so "fixed" comes with proof, and test-driven-development for writing the test before the code.

How it chains with other library skills

Watch-outs

  • Licence: MIT. Free to use.
  • Customer data. A form like this collects names, phone numbers and messages. Test it with made-up enquiries, keep real customer records out of your prompts, and give your OK before it connects to live storage or starts accepting real submissions.
  • Cost: plans and per-task reviews use a lot of tokens on a big feature. If you are on a tight plan, ask for the inline option (executing-plans) instead of one helper agent per task.
  • Long runs: the agent can work for a long stretch on its own once you say go. Keep your work in git, and read the plan before you approve it.
  • Anything public stays yours. Merging to your main branch, pushing and deploying should wait for your OK. Say so in your project rules if your agent is allowed to run commands on its own.
  • Telemetry: the optional visual companion in brainstorming loads a logo from Prime Radiant's website, which tells them the Superpowers version you use. The README says it sends nothing about your project. To turn it off, set the environment variable SUPERPOWERS_DISABLE_TELEMETRY to true.
  • When it misbehaves: ask the agent to "figure out what went wrong with superpowers in this session". A diagnosing skill reads the session and reports what happened.

github.com/obra/superpowers, by Jesse Vincent and Prime Radiant, MIT licence. This page is our own summary; the README is the source of truth for install steps.

MORE LIKE THIS

SKILL · Vibe codingfrontend-design: Anthropic's skill for pages that don't look like a templateREPO · Vibe codingImpeccable: design commands and an AI-look checker for your coding agentREPO · Vibe codingagent-browser: let your coding agent open, click and screenshot your pages