Claude Skills

What is it

Claude Skills are a markdown-based extension system built into Claude Code that lets you encode repeatable workflows as AI-executable instructions. Each skill is a SKILL.md file containing YAML frontmatter and natural language instructions; it becomes a /slash-command you can invoke directly, or that Claude loads automatically when relevant. Skills can include supporting files (templates, reference docs, scripts), be scoped globally or per-project, and chain together into meta-skills. They follow the Agent Skills open standard, which means they work not only in Claude Code but also in Claude.ai and Claude Desktop.

My Opinion

Skills are the most important primitive Claude Code ships. Not the agentic execution, not the MCP integration — the skills. They’re what makes the whole thing compound over time.

Writing Instructions, Not Code

The key insight is that you’re not writing code — you’re writing instructions. The same way you’d onboard a senior colleague to your workflow, except this colleague has perfect recall and never finds repetitive tasks tedious. A SKILL.md file is an SOP for an AI: describe the steps, specify when to ask for user input, encode your edge-case logic, and you’re done. No DSL to learn, no framework to configure. Just markdown.

This shifts the abstraction entirely. Instead of automating workflows by writing scripts, you document them. The documentation is the automation.

The Composable System

Where skills really click is when you make them composable. Each skill does one thing well — scaffold a post, write a concept, generate a cover image — and a thin meta-skill chains them into a full pipeline. My /publish-post skill is about 80 lines of pure orchestration: it knows nothing itself, it just calls the sub-skills in sequence. The practical payoff: fixing a bug in /init-post fixes it everywhere, whether you invoke it directly or through the pipeline.

One gotcha I learned the hard way: disable-model-invocation: true blocks programmatic invocation too, which means meta-skills can’t call sub-skills that have it set. The fix is to write specific descriptions as guardrails instead — Claude won’t randomly scaffold a blog post, but the meta-skill can still invoke it programmatically.

The New Standard

The ecosystem around skills is growing fast. Claude Code has a built-in marketplace. skills.sh and others are emerging as community hubs. More importantly, the Agent Skills open standard is being adopted beyond Anthropic: GitHub Copilot is widely rolling it out, OpenCode supports it natively, JetBrains AI Assistant is moving in that direction. It’s not a Claude-specific format anymore — it’s becoming the interoperability layer for AI workflow instructions across tools, which is exactly what AGENTS.md did for agent context.

Define skills once, use them wherever your team works.

Conclusion

Skills are how you make Claude Code pay off over time. The first skill you write saves you an hour; the tenth compounds into a system that eliminates entire categories of friction. If you use Claude Code seriously, skills aren’t optional — they’re the foundation. Start with the workflow you repeat most often, encode it in plain markdown, and iterate from there.