跪拜 Guibai
← Back to the summary

Superpowers AI Skill Gets Uninstalled: Token Bloat and Over-Planning Kill the Workflow

Hello everyone, I'm Kakarot.

I've been recommending a Skill called Superpowers to everyone.

Over the past few months, this project has gained a lot of stars. The star count has reached 25.2k, which shows how popular it is.

What does this project do?

It's a Skill, meaning it can be used in any AI Agent. For example, your local Claude Code, Codex, Cursor, Trae...

What's the use? 🤔 In one sentence, it sets rules for your AI Agent: Don't start writing code right away.

First, clarify the requirements, do the design, list the plan, write the tests, then implement, check, and verify.

In the ancient programming era, this was a very standard software development process.

So, for beginners, this skill is very friendly.

For example 🌰

Without this skill, your Claude Code or Codex might operate like this:

User: Help me build a login feature
AI: Okay, here's the code...

That is, you raise a question, and the AI directly follows its own plan, imagines what the login process should look like, and then directly modifies the code to complete the development.

This heavily depends on how complete your description is. Frankly, it relies heavily on experience. First, you need to have a clear description of the login feature.

If you have the Superpowers skill installed, it might ask you many questions before starting development.

For example:

A. Is registration required before login?
B. What information is needed for login? Is it directly via username and password?
C. Additional...

When asking questions, it gives you options like A, B, C, and you can also add your own.

After multiple rounds of questioning and confirmation, your Agent knows what you really want to do and clarifies your actual needs and boundaries.

Honestly, for many people, especially non-programmers and non-product managers, being able to clearly describe one's own needs is itself a "technical skill".

This relies heavily on experience. If it's not discussed clearly, the AI directly modifies the code, and there's a high probability it's doing useless work.

Superpowers has many built-in Skills, and it strictly follows the workflow below to help you execute.

1. First confirm what you really want to do
2. Sort out requirements and boundaries
3. Provide a design plan
4. Break down into executable tasks
5. Write tests
6. Implement code
7. Call sub-Agents for review
8. Announce completion only after verifying results

This workflow is very helpful for people who have no experience with code development or for programmers with insufficient experience!

If you are also just starting to learn AI programming, I still strongly recommend you learn about it and use it.

If you're interested, leave a comment, and I can write more in-depth about how to use this skill later 🤔

However, when you become more familiar with the project, and the project gradually becomes complex and the code repository grows larger, the situation changes.

This Skill becomes a bit redundant, even somewhat superfluous.

This is mainly reflected in three aspects:

  1. Superpowers consumes a lot of Tokens.
  2. Context pollution.
  3. Over-planning for small tasks: Small tasks might also generate Spec and Plan documents, and then create a Worktree locally. Every time, I had to tell it to modify directly on the local project branch and not use Worktree.

1. High Token Consumption

I wonder if you've noticed that Superpowers is very Token-intensive. This is also the biggest problem people have always criticized Superpowers for.

Because Superpowers essentially uses a series of Skills to ensure the Agent executes according to a specific workflow. The underlying mechanism is a series of super-long instruction documents to constrain the Agent.

Its main built-in Skills include:

Every single Skill is a super-long instruction document.

For example, the Skill responsible for clarifying our requirements is actually the brainstorming Skill underneath.

When executing the brainstorming process, it might also read:

Therefore, a complete Superpowers workflow can easily consume thousands or even tens of thousands of Tokens in extra context.

Of course, besides brainstorming, there are also Skills for writing requirement documents and execution plans...

Each Skill occupies many Tokens, all of which are loaded into the context.

Because when we converse with AI, the full context is brought along every time.

Suppose you have 10 rounds of conversation in one session. The context occupied by these Skills alone might be repeatedly included 10 times. You can imagine the Token consumption is huge!

2. Context Pollution

Anthropic's official documentation clearly states that as the number of Tokens in the context increases, the model's accuracy and information recall ability may decline. This phenomenon is called "context rot".

System prompts, tool definitions, file content, tool results, and thinking content all occupy context.

When the context is stuffed with:

The truly important business requirements and code details may be diluted, and attention is also dispersed.

Thus, an ironic result emerges:

The Skill was installed to make it smarter, but in the end, the Skill filled up its working memory.

3. Over-Planning

Superpowers automatically triggers brainstorming and full planning for very simple tasks.

For example, some feedback includes:

Some long-term users give very consistent evaluations:

It's great for major changes and refactoring; it's terrible for small changes and trivial tasks.

For example, here, I asked it to do a very simple task. Just extract the links from two sub-pages and put them in the navigation bar.

Here, it used the brainstorming Skill to recommend a solution, which was fine.

But then it started writing a plan document, which I felt was unnecessary, as this task was very simple.

After writing the plan, it went to develop in a Git Worktree. Then I had to ask it to merge the code there back into the current branch.

It complicated a simple matter, really unnecessary 🥹.

Recommended Approach

Since Superpowers is so bloated, what should be done?

Today's large models are far smarter than you imagine. Their planning and execution capabilities are already very strong.

If you are preparing to develop a new project, then the entire Superpowers workflow is quite suitable.

It can help you clarify requirements and then generate an MVP (Minimum Viable Product).

If you are maintaining an old project with many code files, then installing Superpowers is not very necessary. You can just uninstall Superpowers.

If you feel uneasy, or are already accustomed to this workflow, you can just keep these two Skills:

Uninstalling?

How to uninstall in Codex?

Search for super on the plugin page, then click on Superpowers below to open it.

There are toggles here to directly manage each Skill. Just turn on the Brainstorming and Writing Plans Skills.

How to uninstall in Claude Code?

In Claude Code, Superpowers is generally installed as a plugin.

It has over a dozen Skills built-in, and it seems they cannot be individually turned off. If you have a good solution, let's discuss it in the comments.

After opening the skill list via the skill command, you'll find the corresponding Skills are locked and cannot be individually turned off 🤔

Then you can only temporarily disable the entire plugin.

In Claude Code, use the /plugin command to open the plugin system.

Then select the Installed panel, search for super, and you can see the Superpowers plugin.

Press Enter to select it.

Then select Disable Plugin to turn off the plugin. When you need to re-enable it, follow the same operation path.

I'm Kakarot, continuously sharing AI information that is useful to you~