Agent skills are becoming composable the way npm packages did
I made a skill that creates skills, then allmd to turn the universe into a skill. Then the pattern clicked: agent skills are composable units of capability the way npm packages are composable units of code.
Based on building agent skills, shareful.ai, and allmd in early 2026.
An agent skill is a markdown file: instructions in plain language, saved as .md, and your agent knows how to do something new. No API, no SDK, no build step.
Distribution is the interesting part. npx skills add mblode/agent-skills, one command, and it’s in your project. That’s npm install for agent capabilities.
I’ve built skills for Claude Code since late 2025, pulling out any instruction I repeat across projects: deployment patterns, review checklists, animation guidelines, into shareable files.
The meta moment
Then I built agent-skills-creator: a skill that creates skills. Then allmd, which turns any URL, PDF, video or audio into clean markdown, tagline “turn the whole universe into a skill.” Then shareful.ai, Stack Overflow for AI coding agents, because the problems we solve with agents aren’t the ones we Google.
Each felt standalone. They’re one idea: build the ecosystem that makes agent capabilities shareable.
The npm analogy
In 2010, npm was a weird idea. Why install left-pad instead of writing four lines yourself? Because small, focused units compound into capability no single developer could build alone.
Skills are on the same trajectory. One for database migrations, one for accessibility audits, one for performance profiling, and your agent is competent across domains you aren’t expert in.
The early npm ecosystem was messy and exciting. Skills are that now: no dominant registry, best practices still forming.
MCPs are the other half. They’re infrastructure for agents rather than humans: natural language APIs so bots can run commands. Stainless auto-generates SDKs and MCPs from OpenAPI specs, and that layer is standardising fast.
Skills sit on top. The MCP provides the capability, the skill provides the judgment.
The tension
What I keep wrestling with: even skills for CLI tools are mostly overkill, because Claude Code can figure it out. Give it a well-documented tool and it reads the docs and uses it correctly.
So why do they matter? Quality and consistency. The gap between “the agent can do it” and “the agent does it the way I want” is real, and skills close it.
Maybe skills are transitional: essential now, obsolete later. A bet worth making, even if the agents eventually eat it all.