Claude Skills: Stop Re-Teaching the Same Task

The first time I wrote a Claude Skill it was out of pure annoyance. I had pasted the same three paragraphs of "here is how we format a release note" into a chat for the fourth time that month. A Skill is the fix for exactly that feeling: you write the instructions down once, and Claude loads them on its own when the task comes up.

What a Skill actually is

Strip away the framing and a Skill is just a folder with a SKILL.md file inside it. The file has a short header — a name and a description of when to use it — followed by the instructions, any examples, and scripts the task needs. Claude reads the description, decides the task matches, and pulls in the rest only then. That last part is the point: you are not bloating every conversation with instructions you rarely need. They stay dormant until they are relevant.

The description is the whole game. If it is vague, Claude never reaches for the Skill at the right moment. Write it the way you would brief a new hire — "use this when someone asks you to X" — a concrete trigger, not a topic label. Anthropic's official Agent Skills repository is full of working examples — open one and copy the shape before you write your own.

When it is worth it

My rule of thumb: write a Skill the third time you explain the same thing. Once is a fluke, twice a coincidence, three times a pattern — and patterns are what you encode. Good candidates have a fixed shape and a variable input: formatting a doc to a house style, turning a data export into a standard report, running a checklist before you ship.

What one looks like

Here is a whole Skill — the release-note one that started it all. The description is the trigger; the body is what you stop retyping:

---
name: release-notes
description: Use when someone asks to write or format a release note,
  changelog entry, or "what shipped" update from a list of merged changes.
---

Turn the raw change list into a release note in our house style.
Group items under Added / Fixed / Changed, one plain-language line each
(what the user notices, not the commit message). Lead with the biggest
change. End with an upgrade note only if something breaks.

What it is not

A Skill is not a dumping ground for everything you know. Genuinely novel, one-off problems do not belong in one — you would spend longer maintaining the Skill than doing the work. Reserve them for the boring, repeatable jobs that quietly drain a few minutes every week.

This week

Open your chat history and find the instruction you have pasted most often. That paragraph is your first Skill. Move it into a SKILL.md, write a sharp one-line description of when it applies, and never type it by hand again.

Sources

• Anthropic — Agent Skills repository
• Anthropic — skill-creator SKILL.md