跪拜 Guibai
← All articles
Frontend · AI Programming · Claude

Five Claude Code Skills That Actually Change How AI Works, Filtered from 1,400

By kyriewen ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The Claude Code Skills ecosystem is ballooning like VS Code extensions did in 2020 — quantity is high, quality is low. A clear filtering heuristic (does it change workflow or just repackage a prompt?) saves developers from installing noise and directs attention to tools that genuinely reduce rework, token costs, and onboarding time.

Summary

After installing dozens of Claude Code Skills, the pattern is clear: the majority wrap a simple prompt inside a SKILL.md file. Real value comes from Skills that change the AI's behavior, not its instructions. Superpowers (245K stars) transforms the AI from a code-typing assistant into a colleague that asks clarifying questions, writes tests first, and debugs systematically — cutting token usage by 14% in controlled tests. Taste Skill (51.8K stars) injects design rules like spacing multiples and color hierarchy into every UI element the AI generates, curing the generic template look. Graphify (41.8K stars) scans an entire codebase and produces a visual knowledge graph of dependencies and call chains, collapsing weeks of code-reading into minutes.

Two built-in capabilities round out the toolkit: Deep Research performs multi-angle, cross-validated research with cited sources for tech decisions, and find-skills provides a keyword search across the growing ecosystem. The author's filtering heuristic is simple — if you can express a Skill's core instruction in one sentence to the AI, you don't need the Skill. The ones worth keeping alter the AI's thinking process itself.

Takeaways
Most Claude Code Skills are just prompts wrapped in a SKILL.md file; you can achieve the same result by saying the instruction directly to the AI.
Superpowers changes AI behavior from a code typist to a thoughtful collaborator that asks clarifying questions, writes tests first, and debugs systematically.
In controlled tests across 12 tasks of equal complexity, Superpowers reduced token usage by an average of 14% because upfront clarification prevented rework.
Taste Skill injects design rules — 4/8/16/24px spacing multiples, primary-secondary-neutral color hierarchy, consistent border radii — into every UI element the AI generates.
Graphify scans an entire codebase and produces a visual knowledge graph of module dependencies, call chains, and utility references, cutting project onboarding from weeks to minutes.
Deep Research performs multi-angle fan searches, cross-validates sources, flags contradictions with confidence levels, and outputs a structured report with citations.
find-skills provides a keyword search interface across the 1,400+ Skills ecosystem, avoiding manual GitHub browsing.
The filtering heuristic for any Skill: if its core instruction fits in one sentence you can say to the AI, the Skill is unnecessary.
Superpowers, Taste Skill, and Graphify are compatible with Cursor, Codex, and Gemini CLI, not just Claude Code.
Conclusions

The 14% token reduction from Superpowers is counter-intuitive — spending more tokens upfront on clarification reduces total cost by avoiding rework, which inverts the instinct to minimize initial prompting.

The Skills ecosystem mirrors the VS Code extension marketplace circa 2020: rapid growth, low average quality, and a need for curation heuristics rather than bulk installation.

A Skill's value is not in what it tells the AI but in what it constrains the AI to do — systematic debugging enforces a reproduce-locate-verify-fix-test chain that a simple prompt cannot guarantee.

Design taste encoded as rules (spacing multiples, color hierarchy) represents a category of Skill that provides constraint injection rather than instruction replacement, a pattern likely to spread to other domains like accessibility and performance budgets.

Concepts & terms
Claude Code Skills
Modular add-ons for Claude Code that extend the AI's capabilities. Each Skill contains a SKILL.md file with instructions, and optionally scripts or resources. They can be installed via `npx skills add` from GitHub repositories.
Knowledge graph (in codebase context)
A visual representation of a codebase's structure showing modules, their dependencies, call chains, and utility function references. Graphify generates these automatically by scanning the code, enabling rapid understanding of unfamiliar projects.
TDD (Test-Driven Development) sub-module
A workflow enforced by Superpowers where the AI writes test cases before implementation code, then runs the tests to confirm correctness. This is a behavioral constraint, not just a prompt instruction.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗