AI column

Coding@donghyeok99Staff

Superpowers: agent skills that ask and plan before coding

The plugin that makes coding agents ask before they code has passed 300,000 GitHub stars

One line asking for a todo app: the agent asks first, writes a design and a plan, then subagents turn each task green and the branch merges.
One line asking for a todo app: the agent asks first, writes a design and a plan, then subagents turn each task green and the branch merges.
Who
Made by Jesse Vincent and the folks at Prime Radiant, for people who build software with coding agents.
When
Released on October 9, 2025; the latest release is v7.0.1 from October 21, 2026.
Where
Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, OpenCode and more: the README gives install steps for 16 agents. Runs on your machine.

What was hard

Ask an agent to build an app and it often starts writing code without a single question. The Superpowers team noted in a release that a session designing an iOS app never asked about the platform and built a web app.

What it does about it

With Superpowers installed, the agent answers a request to build something with questions instead of code. It plays back what it heard, lets you pick one of two or three approaches, and writes a design doc for you to approve. The approved design becomes a plan of 2 to 5 minute tasks, and each task goes to a fresh subagent that writes the test first and gets reviewed. The skills trigger on their own, so there is nothing extra to type.

What people made with it

gautamkhorana.comGautam Khorana2026-05-06

Built the WordPress Stack Advisor, a free tool that reads a site's CMS and plugins from a URL and returns a stack recommendation, in about 8 hours of session time.

a free tool that takes a URL, fetches the site, detects the CMS and 30+ plugins, and returns a Claude-generated stack recommendation, in roughly 8 working hours of session time
martianlee.github.ioSeongHwa Lee2026-04-18

Started the knowledge base on their blog with Superpowers, and used it to brainstorm another game concept for a Godot project.

The knowledgebase feature I recently added to this blog started with Superpowers, and I also used it to brainstorm a different game concept for a Godot project I am currently working on.

How it runs inside

  1. 1

    hooks/session-start → Claude Code

    Injects the full using-superpowers skillOn startup, clear and compact, the hook reads SKILL.md into the agent's context.hooks/session-start
  2. 2

    You → Claude Code

    One line like "Let's build X"Its rule says a request to build something goes to brainstorming before anything else.skills/using-superpowers/SKILL.md
  3. 3

    Claude Code → brainstorming

    Invokes the brainstorming skillBefore answering or reading files, it loads any skill that applies.skills/using-superpowers/SKILL.md
  4. 4

    brainstorming → You

    One question, 2-3 approaches, a design docIt asks one question at a time and writes the chosen design to docs/superpowers/specs/ for approval.skills/brainstorming/SKILL.md
  5. 5

    You → Claude Code

    Approval → docs/superpowers/plans/…mdwriting-plans writes tasks with file paths, tests and checks.skills/writing-plans/SKILL.md
  6. 6

    Claude Code → Implementer subagent

    A fresh subagent per taskIt sends only what the task needs, never the session history, and reviews spec and quality after.skills/subagent-driven-development/SKILL.md
  7. 7

    Implementer subagent → Claude Code

    Failing test → minimal code → commitWork follows RED-GREEN-REFACTOR, and code written before its test is deleted.skills/test-driven-development/SKILL.md
  8. 8

    Claude Code → Branch with passing tests

    merge / PR / keep / discardIt checks the tests, lets you pick how to finish and cleans up the worktree.skills/finishing-a-development-branch/SKILL.md

What goes in, what comes out

Let's make a react todo listbrainstorming's questions come before any code
Fix this bugThe systematic-debugging skill runs first
figure out what went wrong with superpowers in this sessiondiagnosing-superpowers reads the transcript and reports what went wrong, with evidence

Is it for you

Good for

  • Your agent keeps building the wrong thing and you ask again
  • You want a design doc and tests left behind with each feature
  • You want to plan once and let the agent work on its own for a long stretch
  • You use several agents (Claude Code, Codex, Cursor) and want one way of working

Not for

  • Install it separately for each agent (per the README).
  • The v7.0.1 notes say v7.0.0's questions got annoying fast; they were retuned and the team expects to keep tuning them.
  • brainstorming's visual companion loads its logo from the makers' site with the version by default; set SUPERPOWERS_DISABLE_TELEMETRY to turn it off.
  • On Windows the hooks run through Git Bash, so Git for Windows is needed.

Reviewed by Devalley’s editor before publishing. How the column is made

Sign in to post, comment and like.Sign in

Report

Reason

Read the community rules