In short A skill is a process packaged once: instructions (and sometimes files and scripts) that tell Claude exactly how to carry out your recurring task. Write a skill and from then on Claude does the work your way, the same each time, without re-explaining the steps.
Why skills
A project gives Claude context (what to know), a prompt the task here and now. But when the same process repeats — "format the report by these rules," "check the code against our checklist" — explaining it each time is tiring and the result drifts. A skill fixes the process once.
Analogy: a standard operating procedure for a colleague. Instead of explaining each time how you want the report, you write the SOP once — and they follow it identically.
What's inside a skill
- Instructions — step-by-step "how to do it": tone, format, rules, what to avoid.
- Sample files (optional) — templates, examples of good results.
- Scripts/code (in advanced ones) — where part of the work is deterministic.
- A trigger description — when to apply this skill; from it Claude recognises a task as a fit.
How skills look across products
- Chat / Cowork (Claude 101, Introduction to Claude Cowork) — a skill teaches Claude your process; in Cowork skills are bundled into plugins and shared with a team through the org marketplace.
- Claude Platform (API, Claude Platform 101) — a skill packages a procedure once and is reused across the agent's calls.
- Claude Code (Claude Code 101, Claude Code setup) — skills for recurring engineering tasks.
The idea is the same everywhere: a process as a reusable object.
Validating a skill (evals)
Before relying on a skill (especially in a team plugin), you validate it — run it on examples and see whether it reliably gives the wanted result. The same principle as evaluating a model in Claude API basics: not "by eye," but on real examples.
When to write a skill
- The task repeats, and the result must be the same each time.
- The process is non-trivial (several rules/steps) you'd rather not re-explain.
- You want to share the way of working with a team.
A one-off task needs no skill — a good prompt is enough (prompting and AI Fluency).
Related
The source is the Claude 101, Claude Platform 101, Claude Code 101 and Introduction to Claude Cowork courses. Related: Claude projects and artifacts, MCP.