跪拜 Guibai
← All articles
Artificial Intelligence · Developers

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

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

Agent scaffolding that burns context and slows iteration is now a net negative for capable models like Fable 5. Teams still running heavy skill suites are paying token costs for process guardrails the model already internalized, and the right default is shifting from 'install everything' to 'enable nothing, add only what breaks without it.'

Summary

Superpowers, a Claude Code plugin with over 250,000 GitHub stars, forces every task through a rigid six-step workflow of brainstorming, spec-writing, planning, TDD, coding, and review. What once solved a real problem—reckless agents that coded without thinking—has become a liability. Users report that trivial fixes like renaming a variable balloon from ten seconds to five minutes, while startup alone burns 22,000 tokens before any work begins.

The backlash isn't just about speed. Forcing 14 skill modules into a limited context window divides the model's attention, and some developers find Claude makes more mistakes with Superpowers installed than without. A competing project, fable-skills, takes the opposite approach with only six lightweight nudges at key decision points, and stress testing with Opus 4.8 suggests that's all a modern agent needs.

The underlying shift is model capability. Fable 5 investigates dependencies, plans approaches, and verifies its own output without external prompting. The pattern mirrors early smartphone app ecosystems: flashlight and calculator apps vanished once the OS absorbed those functions. Skill plugins that once compensated for weak reasoning are now redundant overhead, and the correct posture for late 2026 is a well-written CLAUDE.md plus one or two on-demand skills for specific scenarios.

Takeaways
Superpowers injects 14 skill modules that force every task through brainstorm, spec, plan, TDD, code, and review steps.
Startup alone consumes 22,000 tokens—11% of the context window—before any real work begins.
A ten-second variable rename took five minutes with the full workflow enforced.
One user exhausted their entire token quota in five minutes on a simple task.
Some developers report Claude makes more mistakes with Superpowers installed because the 14 skill rules fragment its attention.
The maintainer cut the codebase from 3,150 lines to 977 lines, a 69% reduction, but the forced-process problem remained.
A lightweight alternative, fable-skills, uses only six skills and provides gentle guidance instead of mandatory workflows.
Fable 5 investigates files, resolves dependencies, plans, and self-verifies without external skill prompting.
WebDevCody publicly uninstalled, citing 20 questions to implement a simple feature and preferring to re-prompt a bad first attempt.
The pattern mirrors smartphone utility apps that disappeared once the OS absorbed flashlight, calculator, and QR-scanning functions.
The recommended setup for late 2026 is a strong CLAUDE.md plus one or two on-demand skills for specific scenarios.
Conclusions

Skill plugins follow the same lifecycle as early smartphone apps: they compensate for missing platform capability, then become dead weight once the platform absorbs those functions.

The token economics of heavy process scaffolding flip from helpful to harmful at a specific model-capability threshold, and Fable 5 appears to have crossed it.

Developer intuition about 'more process equals better output' breaks down when the process itself consumes the context window the model needs to reason well.

Uninstalling a 250k-star tool isn't a rejection of its design; it's a signal that the layer it occupied no longer exists.

Concepts & terms
Superpowers
A Claude Code skill plugin that injects 14 mandatory workflow modules (brainstorm, spec, plan, TDD, code, review) into every agent task, originally designed to compensate for models that coded without planning.
fable-skills
A lightweight alternative skill set with only six modules that provides gentle guidance at key decision points rather than enforcing a full workflow on every task.
CLAUDE.md
A project-level configuration file for Claude Code that specifies coding conventions, project structure, and behavioral rules, increasingly used as a lightweight replacement for heavy skill plugins.
From the discussion

The consensus is that skill plugins are a temporary fix. Native model capabilities are absorbing what these tools do, making external scaffolding redundant. One holdout still uses the plugin for heavy reasoning tasks but expects to drop it once models catch up.

Skill plugins should be project-scoped, not global, and even token-saving libraries will fade as models improve.
Building simple, direct, and cheap AI access is the model provider's responsibility, not the user's.
Complex reasoning tasks still benefit from the plugin, but native agent capabilities are rapidly converging on those features, making uninstallation likely soon.
Featured comments
字节别想得到我

This aligns so perfectly with my thinking. Skills like this only need to be set at the project level. Going forward, even various open-source libraries that save tokens will gradually disappear. This is something the large model providers themselves should be doing. Artificial intelligence is not artificial stupidity; it should let anyone get started easily and directly with the fastest and cheapest AI.

Eyki

I installed it, and I enable it for complex thinking and sorting tasks. I agree with what you said. Many agents are converging these useful skills into native capabilities. Maybe after a while, I'll consider uninstalling it.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗