Superpowers Hit 250K Stars, Then Developers Started Ripping It Out

A GitHub project with over 250,000 stars is being collectively uninstalled by its users.

This project is called Superpowers, the most popular skill plugin in the Claude Code ecosystem, bar none. Once installed, it injects 14 skill modules into your Agent, forcing every task through a complete workflow of brainstorm, spec, plan, TDD, code, and review.

Sounds great, right?

But recently, friends around me have been uninstalling it one after another. When I asked why, the answers were surprisingly consistent: it's too slow, it consumes too many tokens, and today's Agents simply don't need this stuff anymore.

I didn't believe it at first. Until I uninstalled it myself.

🔹 First, let's talk about what Superpowers actually did

When Superpowers first came out, it genuinely solved a real pain point.

Agents back then were relatively "reckless." You'd throw a requirement at them, and they'd start writing code immediately—no planning, no testing—and the result would often be full of bugs. You'd have to make them fix it repeatedly, which was actually very inefficient.

Superpowers' idea was: since the Agent itself lacks self-discipline, give it a set of process constraints. Every task must first brainstorm, then write a spec, then make a plan, then TDD, then write code, and finally review. It was like putting a complete SOP on an impulsive intern.

When it was released in the fall of 2025, the entire developer community was recommending it. GitHub stars soared past 250,000, and it entered Anthropic's official marketplace.

But the problem is, that was the Agent of 2025.

🔹 The cost behind 250,000 stars

I really started thinking about uninstalling because a friend complained to me. He asked Claude Code to fix a tiny bug—changing a variable name from camelCase to snake_case.

Normally, this would take ten seconds.

But with Superpowers installed, it first spent a minute brainstorming, then wrote a spec, then made a plan, and only then started changing the variable name. After the change, it still had to run a review.

A ten-second task took five minutes.

After he told me this, I went to GitHub and browsed the issues, and found that many people felt the same way.

One user specifically ran a test: with Superpowers installed, Claude Code consumed 22,000 tokens just to start up, using 11% of the context window before any work even began. Add in the system's own overhead, and the effective workspace was greatly reduced.

💡

GitHub address: github.com/obra/superpowers/issues/190

There's an even more absurd case: someone said that on a simple task, they hit the entire token quota in 5 minutes. The exact quote was "after 5 minutes i went from 0% to 100% quota reached."

A tool that was supposed to save you effort instead became the biggest resource drain.

🔹 It's not just slow, it might also make the Agent dumber

This is something I didn't expect.

A comment from a Hacker News user left a particularly deep impression on me. He said, "I personally don't like superpowers very much. My boss does. I think Claude makes more mistakes when using superpowers than when not." Even his boss thought it was great, but his own experience was that Claude actually made more mistakes with it installed.

Think about it, and it actually makes sense. The Agent's context window is limited. You've stuffed 14 skills into it, each with a bunch of instructions and rules. The Agent has to juggle these rules while processing your request, and its attention gets divided. It's like making someone recite an operations manual while working—they're actually more prone to errors.

The maintainer also realized this problem. They later did a major optimization, cutting the code for the 14 skills from 3,150 lines down to 977 lines—a 69% reduction.

But after the cuts, the core problem remained: it still forcibly inserts a complete workflow before every task.

🔹 The most interesting comparison to me

A developer named DizzyMii created a project called fable-skills, with only 6 lightweight skills. The approach is the complete opposite of Superpowers: no forced processes, just gentle guidance at key points.

He used Opus 4.8 for multiple rounds of stress testing to polish these 6 skills, and the conclusion was that lightweight guidance is completely sufficient.

You don't need 14; 6 is enough.

What does this tell us? It's not that more skills are better, nor that more complex processes are better. The Agent doesn't need you to impose so many rules on it; it just needs a gentle nudge at key points.

The Agent isn't a disobedient child; it's an adult that needs moderate guidance. You give it an employee handbook, not a supervisor watching 24/7.

🔹 So how much stronger has the Agent actually become?

Honestly, the evolution speed of Agents in the past six months has been visible to the naked eye.

My most direct feeling is that after Fable 5 came out, many things that previously required skills to accomplish, it now just does on its own.

Before, if you didn't tell the Agent "think before you act," it would genuinely just charge in recklessly. Now it's different. Throw a complex requirement at Fable 5, and it will first look at the relevant files on its own, sort out the dependencies, figure out a plan, and only then start making changes. After the changes, it will even run tests to verify on its own.

An article on MCP Directory put it bluntly: "Fable 5 investigates before acting and verifies its own work with less prompting." The model itself already knows how to plan and verify; it doesn't need an external set of processes to teach it.

The article's final suggestion is that if you're using Fable 5, at most keep the brainstorm and review modules; delete everything else.

WebDevCody, a well-known developer on YouTube, also publicly tweeted that he uninstalled it. His exact words were: "I guess I'm just not a fan of having to babysit it the entire time and answer its 20 questions to implement a simple feature. I rather let Claude yolo it and reprompt a bad first iteration."

To implement a simple feature, you have to answer 20 questions. He'd rather let Claude go wild on the first try and re-prompt if it fails, than be tied down by the process.

💡

X post address: x.com/webdevcody/status/2006458368387596696

Honestly, I completely understand this feeling.

🔹 This is actually a very classic pattern

Think back to when smartphones first came out.

Back then, you had to install an app for everything on the iPhone. A flashlight required an app, a calculator required an app, a compass required an app, scanning QR codes required an app, even measuring dimensions required an app.

And later? The system came with them built-in. The flashlight was built-in, the calculator was built-in, the compass was built-in, scanning QR codes could be done with the camera. Those utility apps that once had millions of downloads quietly disappeared.

No one misses them.

AI development tools are going through the exact same thing. Early models weren't capable enough and needed external skills to supplement their abilities, just like early phones lacked features and needed third-party apps to fill the gaps. But as model capabilities improve, these external skills become redundant.

When the underlying capability is strong enough, the patches on top naturally fade away. It's not that the skill did a bad job; it's that it has fulfilled its historical mission.

🔹 But I'm not saying skills are completely useless

I need to be clear about this.

Superpowers still holds value for some people. If you're a beginner just starting to write code with AI, you might indeed need a set of processes to help you build good work habits. If you're in a team and need to unify everyone's development standards and enforce a review process, then Superpowers' approach is also reasonable.

And for some particularly complex, large-scale refactors spanning multiple files and systems, sometimes a lightweight plan skill can indeed help the Agent sort out its thoughts.

The key is that skills should be "enabled on demand," not "installed as a full suite."

Just like you wouldn't install 14 productivity apps on your phone and keep them all running in the background. You turn on the one you need and turn it off when you don't.

Anyway, I think the correct posture for the second half of 2026 is: write your project specs well in CLAUDE.md, attach one or two lightweight skills for specific scenarios when necessary, and leave the rest to the Agent itself.

The Agent has grown up; you have to learn to let go.

🔹 In the end

The evolution of the tool ecosystem is actually the same principle as the natural world. Early species protected themselves with exoskeletons, and later evolved endoskeletons, causing the exoskeletons to degenerate. It's not that exoskeletons were bad; it's that internal capabilities became strong enough to no longer need them.

Superpowers is a good product; it provided real value when Agents weren't mature enough. But the wheels of technology wait for no one. Models are evolving, Agents are evolving, and user needs are evolving.

Capabilities get internalized, and patches fade away.

This isn't anyone's failure; this is progress itself.

Are you still using Superpowers or other skill plugins now? Which ones do you think should stay and which should go? Let's chat in the comments; I read every one. Give a like, hit 'watching,' add a star ⭐, and you'll see the next update first.

Thanks for reading my article. Until next time.

💡

/ Author: Kakarot / For submissions or tips, contact email: [email protected]