Skip to article
KnaccoGuides
Start here

What are Claude skills? How they load, where they run and three real examples

What Claude skills are: folders of instructions that Claude loads when a task calls for them. How they load, where they run, what they cost and three examples.

Purple Knacco character reading a book inside a navy bookend, with three books, a cream arch and a smiling star to its right.
0:00
10:54

Claude skills are folders of instructions, scripts and reference files that Claude loads when a task calls for them. Each skill has a SKILL.md file with a name, a description and the steps to follow, and Claude reads the full instructions only when your request matches the description or you call the skill by name.

This guide explains how that loading works, where skills run and what they cost, and then opens three real skills so that you can see what is inside one. The facts were checked on 18 November 2026 against Anthropic's documentation, the open Agent Skills specification and the files of the skills themselves. The sources are listed at the end.

What a skill is

A skill is a folder. The Agent Skills specification requires one file in it, SKILL.md, and allows scripts, reference documents and assets beside it:

skill-name/
├── SKILL.md
├── scripts/
├── references/
└── assets/

SKILL.md has two parts. The first is a short block of settings between two --- lines. The specification requires two settings: a name of up to 64 characters in lowercase letters, numbers and hyphens, and a description of up to 1,024 characters that says what the skill does and when to use it. The second part is the instructions, written as ordinary text.

A complete skill can be a few lines long. This one was written for this guide, and Claude Code's validator accepted it:

---
name: meeting-notes
description: Turns rough meeting notes into decisions, owners and dates. Use when the user pastes notes from a meeting or asks for a summary of one.
---

List each decision in one sentence. Under it, name the owner and the date.
Put open questions last. Do not add anything that is not in the notes.

Nothing is compiled, and a skill made only of instructions needs no programming. Scripts are optional. When a skill includes them, the AI tool runs them and reads only what they print.

The format began at Anthropic and is now an open standard. Claude Code, Codex and Antigravity CLI all load skills from folders that hold SKILL.md, which is why the terms "Claude skills" and "agent skills" describe the same thing.

How Claude decides to use a skill

Anthropic calls the method progressive disclosure. A skill loads in three stages, so that a large library costs little until part of it is used.

Stage What is loaded When
Discovery The name and description of every installed skill At the start of each session
Activation The full text of one SKILL.md When your request matches its description, or when you call it by name
Execution Reference files and scripts in the folder Only when the instructions call for them

Anthropic's documentation puts the first stage at about 100 tokens for each skill and the instructions at under 5,000 tokens. Reference files cost nothing until they are read, and a script's code never enters the conversation, only its output.

Two things follow from this. The description decides whether a skill is ever used, because it is all Claude sees before it chooses. And a skill is not always on. Custom instructions apply to every conversation and a project's files are loaded in every chat inside that project, while a skill's instructions arrive only when they are needed.

You can also skip the matching. In Claude Code, typing /skill-name at the start of a message runs that skill directly.

Three real skills, opened

Anthropic's pdf skill: instructions with scripts

The pdf skill sits in Anthropic's public skills repository. Its folder held 12 files on 18 November 2026: a SKILL.md of 314 lines, two further instruction files named FORMS.md and REFERENCE.md, a license file with the Apache License 2.0, and eight Python scripts, each named for one job, such as fill_fillable_fields.py and convert_pdf_to_images.py.

Its description is a list of the moments when it applies: reading or extracting text and tables, merging and splitting files, rotating pages, adding watermarks, filling forms, and reading scanned pages. It ends with a plain rule: if the user mentions a PDF file or asks for one, use this skill. The length is deliberate, because the description is all Claude matches against. The scripts handle the steps where a fixed program is more reliable than written advice, such as filling the fields of a form.

Knacco's Prompt Writing Guide: a method with a checker

The free Prompt Writing Guide is a folder named kit-prompt-writing-system. It holds SKILL.md, a longer GUIDE.md, three reference files and two Python tools. The method builds a prompt from four parts, which are a role, the input, the task and an example, and one tool checks the result against 14 rules.

This is a real run from 18 November 2026, in Claude Code 2.1.326. The request was of a kind that people type every day, saved as a brief in the skill's four parts:

Role:    You are a world-class copywriter.
Input:   We sell handmade ceramic mugs online.
Task:    Write a great product description for our new mug.
Example: The description should be warm, friendly and persuasive.

Called by name and asked to check the brief, Claude ran the skill's checker. The checker returned a score of 36 out of 100 and reported 10 of its 14 rules broken. Four of its findings:

Part What the checker reported
Role "World-class" praises and does not specify. No level of experience is given.
Input No situation, no constraint and no statement of what is missing
Task No deliverable, no length, no register and no stated use
Example Eight words that describe the output and do not show it

The same request, sent in a fresh session without naming the skill, received general advice and no score, because Claude answered without loading the skill. Both sessions had 96 skills installed. The comparison shows what a skill adds: a fixed method, and in this case a result you can check.

A skill of eleven lines from Claude Code's documentation

Claude Code's documentation builds a first skill named summarize-changes. Its file has a description, one short paragraph of instructions, and this line between them:

!`git diff HEAD`

Claude Code runs that command and puts its output in place of the line before Claude reads the skill, so the instructions arrive with your current changes already included. The feature belongs to Claude Code. Anthropic's documentation says it does not work in the Claude apps or through the API, which is a reminder that a skill written for one tool may do less in another.

Where skills run

Place How a skill gets there What to know
Claude apps Anthropic's skills for Excel, Word, PowerPoint and PDF files are built in. You upload your own as a ZIP file under Customize > Skills. Code execution must be switched on. The Help Center lists skills on the Free, Pro, Max, Ultra, Team and Enterprise plans, while one documentation page names only the paid plans for uploads.
Claude Code A folder in ~/.claude/skills/ or in .claude/skills/ inside a project, or a plugin that carries skills Nothing is uploaded. The folder is read from your computer.
Claude API An upload through the Skills API, used with the code execution tool Skills run in a container with no network access.
Codex A folder in ~/.agents/skills/ Follows the open standard.
Antigravity CLI A folder in ~/.gemini/config/skills/, or in .agents/skills/ inside a project Google's replacement for Gemini CLI.

Skills do not move between these places by themselves. A folder you copy into Claude Code is not uploaded to your Claude account, and a skill uploaded in the Claude apps is not available through the API. The one bridge runs in one direction: skills switched on in your Claude account also load in Claude Code, version 2.1.273 or later, when you sign in with the same account.

How to install Claude skills gives the folders, the upload steps and the commands for each tool.

What a skill is not

  • A prompt. A prompt is text you send once. A skill is saved, and it loads without being pasted. Prompt vs skill vs workflow covers that choice.
  • A project. A project holds background that is always loaded in its chats. A skill holds a procedure that loads when a task calls for it, in any conversation.
  • An MCP server or connector. A connector gives Claude access to another system, such as a calendar. A skill gives Claude a way of doing something. Anthropic's Help Center states the division: MCP connections give Claude the tools, and skills teach Claude how to use them well.
  • A plugin. A plugin is a package that can install several skills at once, with connectors and other parts.

Claude skills vs plugins vs MCP servers vs prompts sets the four side by side.

What skills cost, and what to watch

Money. The feature is part of Claude. Many skills are free, including Anthropic's own and the public collections on GitHub, and some publishers sell theirs. Knacco, for example, has 123+ products, some of them free.

Room in the conversation. Every installed skill adds its name and description to each session. In Claude Code the list has a budget of 1% of the model's context window. When the list overflows, Claude Code drops descriptions, starting with the skills you call least, and a skill without its description is rarely chosen.

Risk. A skill can carry scripts, and its instructions run with your files within reach. Anthropic's advice is to use skills only from sources you trust and to read every file in a skill from anyone else. It names data theft and misuse of tools as the risks. Where to find Claude skills has seven checks to run before you install one.

When a skill is the right tool

Anthropic gives a simple test: create a skill when you keep pasting the same instructions, checklist or procedure into the chat. Three questions sort most cases.

  1. Do you repeat the method? If yes, a skill fits. If the task comes up twice a year, a clear prompt is enough.
  2. Is the missing thing knowledge or access? A skill supplies knowledge. If Claude cannot see the data at all, you need a connector first.
  3. Should it apply to everything? A rule for every conversation belongs in your custom instructions or a project. A skill is for one kind of task.

Where to start

Start with one skill and one real task. Knacco's free skills, prompt packs and workflows cost nothing to try and need a free Knacco account. Each has setup files for Claude Code, Codex and Antigravity CLI, and its page states what it requires before you download it. The Claude skills library lists the skills, for writing, brand work, product photos and video prompts. It is not the place to look for coding skills.

Before you install your first skill

  • You can say which repeated task the skill is for.
  • You have read its SKILL.md and every script.
  • You know what it needs: accounts, software and any paid service.
  • You know where it will run: the Claude apps, Claude Code or another tool.
  • You have one real task ready, and you know what a good result looks like.

Common questions

What are Claude skills in simple terms?

Claude skills are folders of instructions that teach Claude how to do one kind of task. Each folder has a SKILL.md file with a name, a description and the steps to follow, and it may also hold scripts and reference files. Claude loads a skill only when a task calls for it.

Are Claude skills free?

Anthropic's Help Center lists skills on the Free, Pro, Max, Ultra, Team and Enterprise plans. Many skills cost nothing, including Anthropic's own and the public collections on GitHub, and some publishers sell theirs. A skill may still need something that costs money, such as an account with another service, so read its requirements first.

What is the difference between a Claude skill and a prompt?

A prompt is text you send once. A skill is a saved folder of instructions that Claude loads by itself when your request matches its description, or when you call it by name. A skill can also carry scripts and reference files, which a prompt cannot.

Do Claude skills work in Codex and Antigravity CLI?

The format is an open standard named Agent Skills. Codex reads skill folders from ~/.agents/skills/ and Antigravity CLI from ~/.gemini/config/skills/, so a skill written to the standard works in those tools as well as in Claude Code. A feature that belongs to one tool, such as the command lines that Claude Code runs before it reads a skill, may not work in another.

Do I need to know how to code to use or write a skill?

No. A skill made only of instructions is a text file with a name, a description and steps in ordinary language. Scripts are optional. When a skill includes scripts, the AI tool runs them for you, though you should read them before you install the skill.

Are Claude skills safe?

A skill can carry scripts, and its instructions run with your files within reach. Anthropic advises using skills only from sources you trust and reading every file of a skill that comes from anyone else. It names data theft and misuse of tools as the risks.

Sources and further reading (11)
Back to all guides
Keep exploring

A good next read.