跪拜 Guibai
← All articles
Frontend · AI Programming · Claude

Claude Code Skills After Three Weeks: 5 Rules That Actually Make Them Work

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

The gap between a useless Claude Code Skill and a precise one is entirely in the configuration the developer writes around it. A five-minute custom SKILL.md that names the project's actual stack and conventions produces more actionable output than any marketplace Skill used out of the box.

Summary

Most developers install Claude Code Skills and get back generic, unhelpful output. The problem is not the Skill but the lack of project context. A SKILL.md that specifies what to check, what conventions the project follows, and what to ignore turns a 20% useful suggestion rate into 80%.

Three well-tuned Skills outperform 20 installed ones because every Skill consumes limited context-window space. A 10-line custom Skill that names the exact tech stack — Zustand, React Query, Tailwind — beats a 500-line generic one that has to guess. The real efficiency unlock comes from chaining Skills into a workflow where each step's output feeds the next, building a continuous context chain instead of cold-starting every call.

CLAUDE.md sets global project rules, Skills define scenario-specific behavior, and .mcp.json connects external tools. Leaving out CLAUDE.md is like handing out a job manual without stating company policy; the Skill has no idea what stack or standards the project uses. Configure all three and Skills at least double in effectiveness.

Takeaways
SKILL.md must specify what to check, project conventions, and what to ignore; a one-line description produces mostly useless advice.
Limit installed Skills to 3-5 because each one consumes context-window space, and too many cause slow, contradictory responses.
A 10-line custom Skill that names the exact tech stack (Zustand, React Query, Tailwind) outperforms a 500-line generic Skill that has to guess.
Chain Skills into workflows so each step's output becomes the next step's input, building a continuous context chain instead of cold-starting every call.
CLAUDE.md sets global project rules (tech stack, coding standards, forbidden patterns); without it, Skills operate blind to project conventions.
Review and adjust Skill configurations weekly based on actual usage rather than installing and forgetting them.
Conclusions

The most counter-intuitive finding is that a developer's own 10-line Skill beats a community Skill with hundreds of lines, because precision comes from eliminating guesswork about the stack, not from covering more generic scenarios.

Skill effectiveness is bottlenecked by context-window economics: more Skills means less attention per Skill, so curation is a hard requirement, not a preference.

The CLAUDE.md + Skills + MCP layering maps cleanly onto how real teams already separate policy, procedure, and tool access; missing any layer breaks the model.

Concepts & terms
Claude Code Skills
Pre-packaged or custom prompt configurations that define how Claude Code should behave in specific development scenarios, such as code review or component scaffolding.
SKILL.md
A markdown configuration file that defines a Skill's behavior, including its trigger conditions, execution steps, and project-specific conventions it should follow.
CLAUDE.md
A project-level configuration file that sets global rules for Claude Code, such as the tech stack, coding standards, and forbidden patterns, analogous to company-wide policy.
.mcp.json
A configuration file that connects Claude Code to external tools and data sources via the Model Context Protocol, such as databases or API documentation.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗