What Is a Claude Skill? How Editing Skills Work in Claude Code

What a Claude skill is

A Claude skill is a folder of written instructions, and sometimes scripts, that Claude opens only when a task calls for it. Anthropic calls these Agent Skills. The one required file is SKILL.md. It starts with a short YAML header that gives the skill a name and a description, followed by plain Markdown instructions: the steps, rules and examples Claude should follow.

---
name: clean-captions
description: Edits talking-head videos in my clean caption style. Use when I ask to cut, caption or render a video in the clean style.
---

# Clean captions
1. Cut silences longer than 0.4 s.
2. Captions: two lines at most, 64 px, lower third.
...

The folder can hold more than that. A reference sheet with detailed rules, a page of examples, or a small Python script that does one job the same way every time. Anthropic’s engineering team compares building a skill to putting together an onboarding guide for a new hire: everything someone needs to do the job your way, organised so they read only the part they need.

That last idea is the core of the design, and Anthropic calls it progressive disclosure. Claude loads a skill in three stages:

• Name and description. Always loaded, at roughly 100 tokens per skill, so Claude knows the skill exists.
• The SKILL.md body. Read only when your request matches the description.
• Extra files and scripts. Opened or run only when the instructions point to them. When Claude runs a script, only its output enters the conversation, not the code.

Because of this, you can keep many skills installed without crowding Claude’s working memory. Skills work in the Claude apps, the Claude API and Claude Code, though each place installs them separately.

How a skill differs from a one-off prompt

A prompt shapes one conversation, while a skill is saved guidance Claude can pick up in any later session. When you paste a prompt, Claude follows it for that chat and then it is gone. Next week you paste it again, and if a line changes by accident, the result drifts.

A skill sits on disk. Claude reads its description at the start of each session and decides by itself when to load it, or you call it by typing / and the skill’s name in Claude Code. Anthropic’s Claude Code docs give a simple test: make a skill when you keep pasting the same instructions, checklist or multi-step procedure into chat.

That makes the choice practical. Use a one-off prompt when you are trying an idea, fixing one clip, or adding a single effect you may never use again. The free effect prompts on this site are written that way on purpose: copy, fill in the blanks, send. If you write your own, how to write a video prompt for Claude covers the basics. Use a skill when the same set of decisions comes up every time you sit down to edit.

The last row is a different kind of thing. An MCP server connects Claude to something outside, while a skill tells Claude how to do a job. The two work well together, and what MCP means for creators covers that side.

Why an editing style suits a skill

An editing style is a fixed set of choices, which is exactly what a skill is built to hold. Think about what “my usual edit” means for a talking-head video: where the captions sit and how big they are, which words get highlighted, how tight the cuts are, when the frame punches in, which colours the graphics use, and what size and frame rate the final file should be.

Written out properly, that is a long brief. Pasting it into every session is tiring, and small differences between pastes become small differences between videos. Put the brief in a skill and it stays the same each time. The parts that must be exact, such as drawing captions in a set font at a set position, can go into a script that Claude runs rather than rewrites, so the output matches from one video to the next.

A skill also keeps detail out of the way until it matters. A long list of layout rules can live in a reference file that Claude opens only when it reaches the layout step.

If you are new to Claude Code, start with the one-time setup prompt and a few single effects before writing your own skill. Once you notice yourself pasting the same notes into every edit, that is the moment to turn them into a skill. AI Creator Library is preparing editing styles packaged as Claude Code skills for its paid plans; they are not released yet.

Where skills live

Where a skill is saved decides which projects, and which people, can use it. In Claude Code, skills are ordinary folders on your computer:

• Personal: ~/.claude/skills/<skill-name>/SKILL.md works in all your projects on that computer.
• Project: .claude/skills/<skill-name>/SKILL.md inside a project folder works only in sessions in that project. Share the folder through version control and your team gets the skill too.
• Plugin: skills that come inside a plugin are named after it, as /plugin-name:skill-name, so they do not clash with yours.
• Organisation: companies can deploy skills to everyone through managed settings.

A personal editing skill with a helper script might look like this:

~/.claude/skills/
└── clean-captions/
    ├── SKILL.md
    ├── layout-rules.md
    └── scripts/
        └── draw_captions.py

Typing /skills in Claude Code lists the skills it can see.

In the Claude apps, you upload a custom skill as a zip file in settings. Availability depends on your plan and on code execution being turned on, and the Claude Help Center keeps the current details. One catch: custom skills do not sync. A skill uploaded to the Claude app is not available in Claude Code or the API, and the reverse is also true, so install it in each place you use it.

How to change a skill

Changing a skill means editing its files, either yourself or by asking Claude to do it. SKILL.md is plain text, so any text editor works. Edit the description to change when Claude reaches for the skill; edit the body to change what Claude does once it has.

The easier route is to ask in plain English: “Update my clean-captions skill so captions sit 10% higher and highlighted words are yellow.” Claude finds the file and makes the edit. Ask it to show you exactly what changed, so you can check the new rule before your next video.

Claude Code watches the skills folders, so an edit takes effect in the same session without a restart. The exception is a skills folder created after the session started, which needs /reload-skills. One detail matters mid-task: once a skill has been loaded into a conversation, Claude does not reread the file on later turns, so your new version applies the next time the skill is loaded.

A few habits keep changes safe:

• Change one thing at a time and test it on a short clip.
• Keep a copy of the working version, or keep the folder under version control, so you can roll back.
• Keep SKILL.md under about 500 lines, as Anthropic’s docs suggest, and move long reference material into separate files.
• If you only want a skill to run when you call it, add disable-model-invocation: true to its header.

Finally, treat a skill from someone else like software you install. A skill can direct Claude to run code, and Anthropic advises using skills only from trusted sources and auditing every bundled file before using one from an unknown source. That goes for paid skills as well. Anthropic also publishes a skill-creator plugin for Claude Code that helps write and test skills, if you want a guided route.