跪拜 Guibai
← All articles
Backend · AI Programming · ChatGPT

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

By 卡卡罗特AI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent skills that enforce process are a double-edged ledger entry. The same prompt padding that guides a novice burns cash and degrades output quality for an experienced developer on a large repo, making the cost-benefit calculation shift sharply as a project matures.

Summary

Superpowers, a 25.2k-star skill pack for AI agents like Claude Code and Cursor, forces a traditional development cycle: clarify requirements, design, plan, write tests, implement, and verify before declaring completion. For newcomers to AI programming, this structure prevents the agent from jumping straight to code and producing useless output.

As projects grow, the same structure becomes a liability. Each of its eight skills—brainstorming, writing plans, TDD, systematic debugging, subagent delegation, code review, pre-completion verification, and Git worktree isolation—loads a lengthy instruction document into the context window. A single workflow can consume thousands of extra tokens, and those documents are re-sent with every turn of a conversation, ballooning costs and causing context rot that dilutes the model's attention to actual business logic.

Small tasks trigger the full heavyweight process: a simple link extraction spawns a brainstorming session, a written plan, and a separate Git worktree that must later be merged back. The recommended fix is to uninstall entirely for mature projects, or strip down to just the brainstorming and writing-plans skills. Codex allows per-skill toggles; Claude Code currently forces an all-or-nothing plugin disable.

Takeaways
Superpowers injects a full software development lifecycle into AI coding agents, preventing them from writing code before clarifying requirements.
A complete Superpowers workflow can consume thousands to tens of thousands of extra tokens per turn, repeated across every conversation round.
Context pollution from skill instructions, plans, and agent summaries crowds out the actual code and requirements the model needs to reason about.
Trivial tasks like renaming a symbol or extracting links trigger brainstorming, plan generation, and Git worktree creation, adding unnecessary steps.
On mature, large codebases, the skill suite is counterproductive; the recommendation is to uninstall it or keep only brainstorming and writing-plans.
Codex supports per-skill toggles; Claude Code requires disabling the entire plugin because individual skills are locked.
Conclusions

Superpowers embodies a tension that will define AI-assisted development: process scaffolding that protects beginners becomes expensive noise for experts. The tool doesn't change; the user's context does.

The token economics are brutal. A skill document loaded once is a fixed cost, but reloaded across ten conversation turns it becomes a multiplier on every API bill, turning a helpful nudge into a recurring tax.

Over-planning for small tasks reveals a missing primitive in current agent frameworks: a task-complexity threshold. Without it, any structured workflow will eventually feel like bureaucracy.

Claude Code's locked skills expose a UX gap. If a plugin bundles multiple behaviors, users need granular control, not just an off switch, to adapt the tool to the task at hand.

Concepts & terms
Context rot
A documented phenomenon where increasing the number of tokens in an LLM's context window degrades its accuracy and information recall, because the model's attention is diluted across too much irrelevant material.
Git Worktree
A Git feature that lets you check out multiple branches simultaneously in separate directories, used by Superpowers to isolate development tasks so they don't interfere with the main working tree.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗