18 min

Skills: Teach AI Your Way of Working

Package a repeat job once, let the model load it only when needed, and stay safe.

Lesson 6 of 11

0 done

In plain English

A skill is a saved way of working. It is a folder with a SKILL.md file. The model loads it in levels, so unused skills cost little. The description decides when it is used. Install only skills you trust.

Skill: A folder of instructions, and optional scripts, for one job.
Progressive loading: Reading more of a skill only when it is needed.
Description: The short text that tells the model when to use a skill.

Picture it

How a skill is found and used

Name and description known

Always loaded

Your request matches

The model picks the skill

Main steps are read

Only now

Extra files and scripts used

Only if needed

A prompt is a one-time instruction. A skill is a saved way of working. Anthropic calls them Agent Skills. A skill is a folder with instructions, and sometimes scripts and examples. The model reads it only when the job needs it. This lesson explains how skills work and why they are clever.

The problem skills solve

You might explain the same job again and again. How to write your weekly report. How to format a client invoice. How to review a contract. Pasting the same long text every time is slow. And a long prompt fills the model's working space. A skill lets you write the guide once.

What is inside a skill

A main file named SKILL.md. It has a name, a short description and the steps.
Optional extra guides, such as a file just for forms.
Optional scripts the model can run, such as a checker.
Optional facts, such as a price list or a template.

Three levels of loading

This is the clever part. The docs call it progressive loading. The model does not read everything at the start.

Level 1 is a name and a short description. Many skills can be installed with little cost.
Level 2 is the main body. The model reads it when your request matches the description.
Level 3 is anything extra. A script's code stays out of the model's view. Only its result comes back.

The description is the trigger

The model decides from the description whether to use a skill. So it must say what the skill does and when to use it.

a-good-description.txt

name: weekly-report
description: Writes our weekly team report from raw notes.
Use when the user asks for a weekly report, status update or Friday summary.

Where skills work

The Claude apps, for custom skills that you add in settings.
The Claude API, with a code tool.
Claude Code, where skills are folders in your project or home folder.

Skills do not move between these places by themselves. You add them to each one. Other tools have similar ideas, such as saved instructions and projects. Check what your tool offers.

Safety comes first

A skill can contain code and instructions. A bad skill can make the model do harmful things. The docs give a clear rule. Use skills only from sources you trust, such as your own work.

Read every file in a skill before you install it.
Be wary of skills that fetch things from the web.
Treat a skill like software you install.

A small example

Imagine you write a weekly team report. Your skill has three parts. The main steps say how to gather notes, group them and write a short summary. A second file holds your report layout. A third file holds a list of team names and what each team does. When you ask for the weekly report, the model reads the main steps. It opens the layout file only when it needs it.

→Start small. One skill, one job, one page of steps. Add extra files only when the main page gets long.

⚠Never install a skill from a stranger without reading it. It can read files and run code with your access.

Key takeaways

✓A skill is a saved way of working, stored as a folder with a SKILL.md file.
✓The model loads it in levels, so unused skills cost almost nothing.
✓The description must say what the skill does and when to use it.
✓Skills do not sync between products, so add them where you need them.
✓Use only skills you trust, and read them first.

Hands-on exercise

Draft your first skill

Pick a job you repeat each week.
Write a name and a two-sentence description with a clear when-to-use line.
Write the steps in plain numbered sentences, with one example output.
List what extra files it might need later, and leave them out for now.

Check yourself

Q1Why do skills load in levels?show answer

So only the needed parts enter the model's working space. Unused skills cost very little.

Q2What decides whether a skill is used?show answer

Its description. The model matches your request against it.

Q3What is the main safety rule for skills?show answer

Use only trusted skills and read them first. A skill can run code and direct the model.

← Prompt Caching: Pay Less for Repeated Text Tools the Provider Runs for You, and MCP →