Superpowers Plugin for Claude Code: Plan, Test, and Review Explained

By Beau Johnsonยทยท7 min read

Quick answer

Superpowers is a free, MIT-licensed GitHub repo by Jesse Vincent that gives AI coding agents like Claude Code, Cursor, and Codex a full workflow: brainstorm first, write a detailed plan, run each task in a fresh subagent, write tests first, and review code between tasks.

In this video I break down Superpowers, an open-source GitHub repo with over 40,000 stars that changes how AI coding agents like Claude Code, Cursor, and Codex write your code. If you've ever watched your agent jump straight into code, skip the tests, and leave you babysitting the mess, this one is for you, especially if you're newer to vibe coding and setting up AI agents.

I recorded this in March 2026. Tools like this change fast, so check the current docs in the repo before you set anything up.

What I cover in this video

  • 0:00 โ€“ The problem: AI coding agents guess, skip tests, and lose the plot on bigger features
  • 1:32 โ€“ What Superpowers is and who built it (Jesse Vincent)
  • 2:18 โ€“ The brainstorming skill that makes the agent stop and ask questions first
  • 4:35 โ€“ Detailed implementation plans with 2โ€“5 minute tasks
  • 5:20 โ€“ Subagent-driven development
  • 6:06 โ€“ Test-driven development and code review between tasks
  • 6:51 โ€“ What other developers are saying about it
  • 7:37 โ€“ Which coding agents it works with, and why this matters for my own agent setup

The problem with AI coding agents

You fire up your agent, tell it what you want, and it just goes. It writes code before it understands the problem, skips the tests, and produces spaghetti. Whenever I asked it to do something complex, anything spanning more than a couple of files, it would forget what it was doing halfway through. I'd spend more time fixing the work than if I'd just done it myself.

Superpowers isn't a prompt template or a quick hack. It's a whole methodology for how AI agents should build software, and it's completely free under an MIT license.

Brainstorming and planning before any code

The first thing Superpowers does is force your agent to stop and think before writing a single line of code. Normally you'd say "Build me this feature" in Claude Code and it immediately starts typing. That burns tokens, especially if you're not using plan mode, and it often produces bad output.

With Superpowers installed, the agent activates a brainstorming skill. It asks you what you're actually trying to build, what your edge cases are, and what happens when things fail. It refines the spec through the conversation and shows it to you in chunks short enough to actually read.

I also recommend that any project starts with a product requirements document. Without one, your agent won't know what you're building.

A developer named Eric Grill wrote that he shipped maybe three finished products but started around 47 others. That wasn't because he couldn't code. He couldn't stop himself from jumping in without thinking. The brainstorming skill fixed that for him, and it helped me too.

After you approve the design, Superpowers writes an implementation plan. In the project's own words, it's detailed enough that "an enthusiastic junior engineer with poor taste, no judgment, and no project context" could follow it:

  • Every task is scoped to two to five minutes
  • Every task has an exact file path
  • Every task has verification steps

Subagents, test-driven development, and code review

Subagent-driven development

Instead of one agent trying to hold your entire feature in its head and burning through context, Superpowers spins up a fresh subagent for every task. Each one gets clean context with exactly the information it needs and nothing else. Richard Porter wrote that features spanning 15+ files, which used to make Claude lose track of earlier decisions, now run without problems. That means fewer regressions, better code, and less babysitting.

Test-driven development

This is the part I think matters most if you're building real things. The agent writes a failing test first, watches it fail, writes the minimal code to make it pass, watches it pass, and then commits. If code gets written before the tests, it deletes it. It won't let you skip that step.

Code review between tasks

Between every task, a code review skill checks the finished work against the original plan. Critical issues block progress until they're resolved. It's like having a senior engineer doing continuous code review on every commit.

What developers are saying, and why it matters for builders

Colin McNamara said his productivity with Superpowers beats what his entire team at Oracle Cloud Infrastructure could produce. On the Claude Code subreddit, someone called it the single best plugin ever created. Someone on a productivity subreddit said that when they tried it on a small project, every phase of development got proper attention for the first time, and the output matched what they had planned.

Jesse Vincent isn't some random developer, either. He's been in open source for decades, built Request Tracker (one of the most-used help desk systems in the world), and runs a company called Prime Radiant.

Superpowers also works with pretty much every coding agent. Claude Code has a one-click plugin install, and it's supported in Cursor, Codex, OpenCode, and Gemini CLI.

Here's why it matters to me. I run 12 AI agents on a Mac Mini that handle my content pipeline, videos, blog posts, and podcast outreach. My biggest challenge early on was getting those agents to be disciplined, to follow a process and not freestyle their way through tasks. Superpowers solves that for coding, and its principles apply to everything we do with agents: systematic over ad hoc, evidence over claims, reduce complexity, and verify before declaring success.

Installing it takes two commands. The exact steps are in the repo.

Key takeaways

  • Superpowers is a free, MIT-licensed methodology that makes AI coding agents brainstorm and plan before writing any code.
  • Its plans break work into 2โ€“5 minute tasks with exact file paths and verification steps.
  • Each task runs in a fresh subagent with clean context, which helps on large, multi-file features.
  • It enforces test-first development and runs a code review between tasks. Critical issues block progress.
  • It works with Claude Code, Cursor, Codex, OpenCode, and Gemini CLI.

FAQ

What is the Superpowers GitHub repo?

Superpowers is an open-source, MIT-licensed project by Jesse Vincent that gives AI coding agents a full development methodology: brainstorming, detailed planning, subagent-driven development, test-driven development, and code review. When I made the video it had over 40,000 GitHub stars.

Which AI coding tools does Superpowers work with?

It works with Claude Code, which has a one-click plugin install, and is also supported in Cursor, Codex, OpenCode, and Gemini CLI.

What is subagent-driven development in Superpowers?

Instead of one agent holding your whole feature in its context, Superpowers spins up a fresh subagent for each task. Each one gets only the information that task needs, which means fewer regressions on large features.

Is Superpowers free?

Yes. It's completely free and open source under an MIT license, and installing it takes two commands.

Chapters

  1. 0:00Intro
  2. 1:30The problem with AI coding agents
  3. 3:00What Superpowers actually does
  4. 5:30Brainstorming and planning phases
  5. 7:00Subagent driven development
  6. 9:00Real developer testimonials
  7. 11:00Why this matters for builders
  8. 12:30How to install it

Transcript

Read the full transcript

0:00A developer just built an AI agent operating system that has over 40,000 stars on GitHub. And I'm going to be breaking it all down on why it's going to be useful for you. That's not a typo. I didn't misspeak. 40,000 GitHub stars proves the proof of concept, and it's called Superpowers, which is a great name, by the way. It will completely change how AI coding agents write your code. But here's the problem. And if you've used Claude code or Cortex or Codium or Courser or whatever, whatever coding tools you use, you already know what I'm about to say. You fire up your agent, you tell it what you want, and it just goes, right? It starts guessing, it writes code before it even understands the problem, it skips the tests, it produces spaghetti that you have to babysit for the next 2 hours. Yay! And

0:47I've been there, and you know, I've been pulling my hair out getting Claude code set up before getting Open Claude set up. You know, you want it to build features, and it just ends up with AI slop in the background. So, you know, every time I'd ask it to do something complex right out the box, you know, something spanning more than just like a couple files, it would just lose the plot. The AI agent would go off eating crayons in the corner somewhere. And I know we've all been there before. It would just basically forget what it was doing halfway through. It would skip the test entirely, and I'd spend more time fixing the work than if it I just done it myself. So, that's exactly what this developer had in mind. How can it help people who are just getting started in this space, you know, getting AI agents set up, getting going vibe coding. And

1:32Jesse Vincent was dealing with this, too, which is the developer at hand here. He built Superpowers. It's going to be linked down below, by the way. Anytime I ever make these videos, check the links below, because that's where all the sources are. And you know, it's not just like a prompt template. It's not a hack to get rich quick. It's basically just like an entire methodology for how AI AI agents should be built for software and is completely open source. MIT license, completely free. I'm going to break down what it actually does because this is where we're going to be headed in the future. I just think it's really cool. That's why I'm sharing it with you. The first thing Superpowers does is it forces your AI agent to stop and think before writing a single line of code. That's

2:18going to help you save a lot on token spend down the road. And I know what that sounds like. It sounds basic, but if you think about how you normally use Claude code, you end up spending a lot of tokens in writing a bunch of code, especially if you're not using plan mode, and you're just wasting tokens. So, you would say like, "Build me this feature." and immediately starts typing, no questions, no planning, just code, and it ends up just producing more often than not a bad output. But with Superpowers installed, your AI agent activates what they call a brainstorming skill. It stops, it asks you questions, what you're actually trying to build. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then

3:05you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full Open Claude course, and four live bootcamps every single week where we actually help you get set up from scratch. Like, actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching so you could book a call with me, we could share screens, and I can help you get Claude code or Open Claude running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below.

3:50Get in now before the price goes up. What your edge cases are, what happens when this fails, it refines the spec through the conversation and shows you in chunks short enough to actually read. I do recommend if you're starting any project, the most necessary file to get going is a product requirements document. Without that, your agent will not know what you're building. So, back to the video. So, that alone, it will save you a ton of time. It would have saved me a ton of time in my early builds. Not just the product requirements document, but this GitHub repo right here that I have linked down below. So, one developer named Eric Grill wrote about a whole blog post about this, linked down below again. He said he shipped maybe three finished products, but started approximately 47

4:35others. Sounds like me. Not because he couldn't code, but because he couldn't stop himself from just jumping in without thinking. The brainstorming skill fixed that for him, and that helped me, too, as well. You know, after you approve the design, Superpowers creates an implementation plan. And this is the key here. The plan is so detailed that this is the This is their actual words right here. An enthusiastic junior engineer with poor taste, no judgment, and no project context could follow it. Very simple. Every task is scoped to two to five minutes. Every task has an exact file path. Every task has verification steps. You know, there's no ambiguity. And then here's where it gets a little bit more interesting. It launches what

5:20they call subagent driven development. So, instead of one agent trying to hold your entire feature in its head and burning through context, Superpowers spins up a fresh subagent for every single task. Each one gets clean context with exactly the information it needs and nothing else. Richard Porter, a developer who wrote about this extensively, said features spanning 15-plus files that previously caused Claude to lose track of earlier decisions now just execute flawlessly. The sub-agent architecture means that tasks get exactly what it means and nothing more. Fewer regressions, better code, less babysitting and resulting in your time saved and better output. Here's the part that I think that matters the most for anyone building

6:06actual real stuff. It enforces a true test driven in development stage. You need to actually think through what you're actually building and this prompts you to do exactly that before you even write a single line of code. Well, the agent writing code, not you. So, it's going to write the failing test first, watch it fail, write the minimal code to make it pass, watch it pass, and then commit. And if you try to write code before tests, it literally deletes it. So, it will not let you skip that step. So, between every task, a code review skill activates and checks the completed work against the original plan. Critical issues block progress until they are resolved. It's like having a senior engineer doing the continuous code review on every single commit. Now, I'm going to tell you why this matters for us. For the builders

6:51out there, you and me. For the people in my community who are actually shipping real things, Colin Macnamara, and this dude is not messing around. He said his productivity with Superpowers exceeds what his entire team's at Oracle Cloud Infrastructure could produce. One person outperforming out outperforming entire teams. That's crazy. So, that's someone who worked at Oracle saying this. And on the Claude code subreddit, someone called it the single best plugin ever created. Said they use it 90% more than anything else. Called the creator a legend. Another person on the productivity subreddit said they tried it over the holidays on a small project and for the first time every phase of development got proper attention. So,

7:37there was no rushing through the step process. No skipping validation, the output actually matched what they had planned. So, who is Jesse Vincent? Huh, he's got some He's not some random developer, some guy on the side of the street. He's been in the open-source game for decades. He built request tracker, which was one of the most used help desk systems in the world. He also runs a company called Prime Radiant. The dude knows what he's doing and has an amazing name. I mean, come on. And what's what what's really cool about Superpowers is it works with everything. Claude Code has one-click plugin install. Cursor even supports it. Codec supports it. Open Code supports it. Even Gemini CLI supports it. So, no matter what coding agent you're using, you can add this. Here's what I want to connect

8:23with um back to like what I'm building and you know, reference that and how it relates to me is like I run 12 AI agents on a Mac Mini. You know, my entire business runs behind it. Content pipeline, videos, blog posts, podcast outreach, everything. It's all automated. And the biggest challenge I had early on getting it set up was getting these agents to be disciplined. That has also a lot to do with the memory as well. So, it was very hard to get them to follow a process, to not just freestyle their way through tasks. That's exactly what Superpowers solves for coding and it's one of the greatest philosophies behind you know, coding principles. It applies to everything we should be doing with agents. Systematic over and ad hoc. Evidence over claims. Complexity reduction. Verify before

9:08declaring success. So, if you're building with Claude Code right now, go install this, please. It takes two commands. That's it. And your coding agent goes from helpful assistant to something that actually can ship a production-ready product. The link is down below, completely free. Like I said, 40,000 developers already figured this out, so don't sleep on it. And if you're building something with AI right now, whether it's a SaaS, an automation, a content pipeline, whatever, if you're looking to get Open Claude installed, we've got 97 builders in Shipping School doing exactly that. We now have five calls per week, real people, real projects, real results. Come build with us. It's so much better building together inside of a community rather than by yourself doing nothing. Come join us. The price is going to go up if we just get three more members and you don't want to be locked out of that tier. So, I'll put the link to that down

9:54below and we'll see you in about an hour when the next video drops. Bye-bye.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading