What Is a Claude Skill? A Plain-English Explainer

Anthropic launched Claude Skills in October 2025 and a lot of the early writing about them is either deep in the docs or assumes you're already a developer. This post explains what they actually are, what they do, and whether they're useful to you — without the jargon.

The one-sentence version

A Claude Skill is a folder of instructions you give to Claude once, and Claude uses those instructions automatically whenever your question matches what they're for.

That's it. Everything below is detail.

A concrete example

You upload a "Warren Buffett Thinking Framework" skill to Claude.ai. Inside that skill, there's a description saying something like:

Use when evaluating investments, capital allocation, or long-term value decisions. Apply Buffett's mental models — moats, margin of safety, owner earnings.

Two weeks later you ask Claude a completely unrelated question: "Should I invest in a SaaS company trading at 40x earnings?"

Claude reads that description, realises your question matches, and invokes the skill automatically. You get a response reasoning through Buffett's specific mental models instead of Claude's default helpful-assistant voice. You didn't type a command. You didn't paste anything. Claude just knew.

That's what makes Claude Skills different from every previous way of "customising" an AI.

What's inside a Skill, technically

A Claude Skill is a folder with this structure at minimum:

my-skill-name/
└── SKILL.md

The SKILL.md file has two parts — YAML frontmatter and markdown content:

---
name: warren-buffett-framework
description: Use when evaluating investments, capital allocation,
  or long-term value decisions. Apply Buffett's mental models.
---

# Warren Buffett Thinking Framework

[the actual instructions Claude will follow]

The description field is the critical part. It's what Claude reads when deciding whether to invoke the skill. A sharp description means the skill fires on the right questions and stays quiet on others. A vague description means it either fires on everything (annoying) or nothing (useless).

Skills can also include supporting files — reference documents, scripts, templates — in subfolders. For thinking frameworks you don't need any of that. Just SKILL.md.

How Claude decides to use a skill

This is where the magic is, and it's also where beginners get confused.

When you ask Claude a question, before answering, Claude scans every skill you've loaded and reads each skill's description field. It matches your question against those descriptions. If one matches clearly, Claude loads that skill's full content and answers with it active.

Three outcomes are possible:

1. Clear match → skill invoked. "Should I invest in this business?" lights up the Buffett skill's description. Skill fires. Answer uses Buffett's mental models.

2. No match → no skill used. "What's the capital of France?" doesn't match any skill. Claude answers normally.

3. Multiple matches → Claude picks best. If you have both a Buffett skill and a Munger skill loaded, and you ask about investing, Claude picks whichever has the tighter matching description. You can also override manually by invoking a specific skill.

You can also force-invoke any skill by typing /skill-name at the start of your message. That works even when Claude's automatic matcher wouldn't have triggered the skill.

Where Claude Skills work

Anthropic ships Skills across their entire product surface:

• Claude.ai (web, desktop, mobile) — upload via Settings → Skills
• Claude Code — unzip to ~/.claude/skills/ for personal or .claude/skills/ for project-specific
• Claude API — upload via the API's Skills endpoint (currently in beta) or just load SKILL.md as the system prompt

Your Claude plan matters. Skills are available on Free, Pro, Max, Team, and Enterprise plans. Claude Code users get them too, as does any app built on the Claude API with code-execution enabled.

What Claude Skills are good at

Persistent expertise. Load a skill once, and it's available for every conversation you have. You're not re-pasting prompts or explaining context. The skill sits in the background and activates when relevant.

Multi-domain stacking. Load five skills, and Claude picks the right one per question. Investment skill for one conversation, product design skill for the next, writing skill for the third. Same Claude, different "modes" based on what you're doing.

Automatic invocation. The big one. Most ways of giving AI a custom personality require you to type something to activate it. Skills just work — Claude reads your question, picks the right skill, applies it.

Team sharing. If you share a project with a team, skills committed to the project's .claude/skills/ folder apply consistently for everyone. No more "but it worked on my machine" for prompt engineering.

What Claude Skills are bad at

Instant, hyper-specific tasks. If you want a one-off answer styled like Yoda for a single Slack message, don't load a skill. Just write it into your prompt. Skills are for repeatable patterns.

Contradictory frameworks. Don't load five skills that point in opposite directions. Claude will try to balance them and produce mush. Pick skills that share a worldview.

Highly dynamic context. Skills are static — written once, used many times. If your task needs live API data or the current date or user-specific information, that goes through tools (MCP) or the conversation itself, not a skill.

Replacing other instruction formats. A skill doesn't replace your Claude Projects (those are for documents), Claude Code configuration files (those are for coding context), or MCP tools (those are for actions). They complement. Think of a skill as how Claude thinks, not what Claude has access to.

Skills vs other ways to customise Claude

Quick honest comparison:

Claude Projects — Upload documents as knowledge. Good for giving Claude a lot of information to reference. But everything in the project is always active, which creates noise. Skills are selectively invoked.

Custom Instructions (the old "About me" / "How to respond" fields) — Short, always-on personality nudges. Character-limited. Skills go much further, much more structured, and can be swapped in and out.

MCP (Model Context Protocol) servers — Tools Claude can call. Completely different purpose. MCP lets Claude do things (send email, query a database). Skills shape how Claude thinks about things. Often complementary — a Notion MCP + a "thinking framework" skill works well together.

Prompting tricks in individual messages — Still valid for one-offs. But when you find yourself typing the same instructions repeatedly, you want a skill.

Who writes skills?

Three groups:

1. Anthropic themselves. They ship official skills (PDF manipulation, frontend design, XLSX work) as part of Claude. Good quality, broadly useful.

2. Developers and teams building workflows. They write skills for their specific codebase, domain, or workflow. Usually stored in repos and shared internally.

3. Third-party skill publishers. Marketplaces like authority.md (us), KissMySkills, and various community sources sell or share skills for specific use cases. Authority.md focuses on thinking frameworks — mental models from documented thinkers like Warren Buffett, Steve Jobs, Richard Feynman. Other marketplaces focus on role-based skills (marketing manager, PM, sales rep) or developer workflows (code review, deployment).

You can also write your own. Anthropic publishes a skill-creator skill that helps Claude guide you through building one.

How to try one right now

If you want to see Skills in action in the next 60 seconds:

1. Go to Claude.ai
2. Settings → Skills → Add skill
3. Upload any valid skill zip
4. Start a new chat and ask a question that matches the skill's purpose

If you don't have a skill file yet, we sell thinking framework skills at authority.md/browse — $4.99 each, includes the native Claude Skill zip plus a plain .md for other AIs. You could download one, install it, and watch Claude auto-invoke the framework in about 90 seconds from purchase to first response.

For a full walkthrough of the install process, see our Claude Skills install guide.

The honest limits worth knowing

Claude Skills are useful, but they're not magic:

• Description writing is a craft. Bad descriptions mean bad auto-invocation. A skill with a vague description may not fire when you need it. This is where most amateur skills fall down.

• Skills don't learn. They execute the same instructions every time. If you want Claude to improve a skill based on your feedback, you need to edit the SKILL.md yourself and re-upload.

• Model quality still matters. A brilliant skill loaded into a weak model produces mediocre output. A brilliant skill loaded into Claude Sonnet produces what Sonnet is capable of, with the skill's nudge.

• Format specification may evolve. Skills are six months old. The SKILL.md format is stable, but edge cases and advanced features (scripts, executable code, nested references) are still shifting. Simple skills are safe; anything fancy is subject to breakage.

What to do if this sounds useful

If your AI use falls into repeatable patterns — investing questions, product decisions, creative writing in a specific voice — Skills are almost certainly worth learning. Spend 20 minutes writing one for your most common task and see what happens.

If your AI use is mostly one-off questions — "help me write this email", "explain this code" — Skills are probably overkill. Stick with good prompts.

If you want thinking frameworks in particular — the mental models and decision patterns of specific thinkers, packaged as ready-to-install Claude Skills — that's exactly what we sell:

Investor

Warren Buffett

Value investing, capital allocation, economic moats

$4.99 · View framework →

---

Got a question about Claude Skills this didn't answer? Email us. If enough people ask the same thing, we'll add it here.