Claude Skills Explained: Codify Your Workflows Into Reusable Playbooks

Skills are a big deal. When Anthropic released them, a set of legal AI SaaS stocks sold off 10–15% in a single day because the market realized that specialist workflows could now live inside Claude as portable text files. If you are a knowledge worker, the opportunity goes the other direction: skills are where your domain expertise becomes durable, reusable, and shareable.

This module covers what a skill actually is, how it differs from prompts and projects, and how to build your first one. We’ll build a simple one together so you can see the mechanics — then point at the harder examples in later modules.

What a skill really is

A skill, for our purposes, is a markdown text file with instructions. That’s it. It sits in a known location, and Claude reads it when relevant. Conceptually it has three parts:

1. Metadata at the top — a short description of what the skill is for. This is what Claude scans when deciding whether a task matches.
2. Instructions — the body of the skill. Detailed rules for how to do the task.
3. Examples (optional) — concrete inputs and outputs that teach Claude the pattern.

More advanced skills can also include scripts for data processing. That’s mostly for engineers. We ignore it for now.

Why “supercharged prompt” is the right mental model (mostly)

If you used AI a lot in 2025, you probably built a prompt library — detailed prompts saved somewhere, pasted in when needed. Skills are that idea, done right.

Three things make skills better than a prompt library:

1. You don’t have to invoke them. Claude reads the skill’s metadata in the background and decides to use it when the task matches. No “please use the XYZ prompt” incantation.
2. Examples are first-class. You can include sample inputs and outputs alongside the instructions. Claude learns the pattern, not just the rules.
3. They are portable. A skill is a text file. Share it, version it, sell it.

Where the prompt analogy breaks down: skills can be composed, they don’t eat context tokens unless Claude actually pulls them in (that’s the “progressive disclosure” design), and they enforce consistency across sessions in a way that a copy-pasted prompt never does.

Skills vs projects

Easy to confuse. Clear distinction:

• A project holds context — the files and reference material for a body of work.
• A skill holds a workflow — the sequence of actions and the output format for a task.

You often use them together. The project has the reference memos; the skill has the formatting rules for the resulting report.

Two kinds of skills

Anthropic’s guidance splits skills into two types:

1. Workflow skills — codify how you do something. This is where 95% of the value is for knowledge workers.
2. Knowledge-gap skills — give Claude access to something it doesn’t know (a niche industry standard, an internal framework). Useful, but less of what we’ll focus on in this course.

Start with workflow skills.

Finding the skills menu

In Cowork:

1. Click Customize (the briefcase icon in the left sidebar).
2. Click Skills.
3. You’ll see two sections: My Skills (empty to start) and Example skills from Anthropic.

The most important built-in skill is the skill-creator. Make sure its checkbox is on. We’re about to use it to build our first skill.

Building your first skill (dad jokes)

We are going to build a deliberately silly skill — a “dad jokes” skill — because the mechanics are identical to a real one and the silliness makes each step easy to see. Once you’ve done this, the jump to a real workflow skill is small.

In a new Cowork chat (Sonnet is fine for this — it’s all text):

I want you to create a skill using the skill-creator skill.
The skill tells corny dad jokes appropriate for kids with
some AI and tech nerd humor. The joke should be added to
the bottom of any chat response.

The phrase “using the skill-creator skill” is important. Say it explicitly. If you don’t, Claude may try to build the skill with its general reasoning instead of invoking the dedicated skill-creator. You can tell it’s working when you see a message like “checking if a dad jokes skill already exists” and “creating dad-joke.md” in the response.

Claude will interview you briefly: tone, length, whether the joke should be before or after the main answer, what kinds of jokes. Answer naturally. When it’s done, it will show you the skill file and offer a Copy to your skills button. Click it.

Open Customize → Skills again. Your new skill should appear under My Skills. Make sure its toggle is on.

Seeing the skill in action

Start a new chat. Ask any unrelated question:

What's the difference between Opus 4.7 and Sonnet?

Watch the response carefully. Claude should answer the question — and then, at the bottom, tell you a dad joke. In the sidebar you’ll see “reading the dad-joke skill” when it fires. If you don’t see that, the skill isn’t loading. Check the toggle.

Notice you didn’t say “use the dad joke skill.” Claude figured it out from the metadata. That’s the whole trick.

Editing a skill in plain English

You don’t have to open the markdown file. To change the skill:

1. Go to Customize → Skills.
2. Click the three dots next to dad-joke.
3. Click Edit with Claude.
4. Type what you want changed:

Also do sports-related jokes and math-related jokes.
Make them slightly edgier but still kid-appropriate.

Claude rewrites the skill. You’ll see a new file preview. Important: click Copy to your skills again — it will ask if you want to replace the existing skill. Say yes. Until you click replace, the old version is still the active one.

The two rules for working with skills

After you’ve done this a few times, two habits matter more than anything else:

1. Always verify the skill is being read. Look for “reading the dad-joke skill” (or whatever) in the sidebar. If you don’t see it, the skill isn’t firing. Check the toggle, check whether you’ve turned off all tool access, check that you actually clicked “replace” when you edited.
2. When you edit, click replace. An edit without a replace is just a draft. The old version is still what runs.

What to build next

Once the dad-joke flow makes sense, think about something you do often that is annoying to re-explain every time. Good starting points:

• A research analyzer — you drop in a research report and it extracts predictions, covered companies, bullish/bearish calls, price targets, worst-case scenarios in a consistent format.
• A company style guide — uploaded examples of your firm’s writing plus rules. Every doc you generate now sounds like you.
• An inbox triage skill — read the last 24 hours of email and classify into reply-needed, FYI, delete-worthy, scheduling.
• A transcript parser — turns a meeting transcript into a structured summary with action items, decisions, and follow-ups.

We cover several of these in later modules. The common thread: a workflow you do often, with a consistent output format, that is currently a pain to explain. Those are the workflows worth codifying.

Common mistakes

• Skipping “use the skill-creator skill.” Claude will try to reason its way to a skill on its own, and you’ll get something mediocre. Say the magic words.
• Not checking the sidebar for “reading the skill.” If your skill doesn’t fire, it’s not being read. Metadata may be off, or the toggle might be off.
• Editing without replacing. The edit is just a preview until you click replace. People build a great v2 and wonder why the output didn’t change.
• Building too broad a skill on day one. “A skill that handles all my analyst work” will not work. Start narrow — one task, one output format.
• Not sharing. If you build a skill that works, send the markdown file to your teammates. That is the whole point of skills being portable text files.

A note on what skills are about to do

Skills are portable, composable, and easy to share. That combination is why they’re going to restructure parts of the specialist software market. The pattern is already visible in legal. Expect it to show up in finance, consulting, compliance, and anywhere else where the value of the software is mostly codified domain expertise.

You don’t need to wait for a marketplace. The skills you build for your own work compound. In six months you’ll have a personal library that makes you measurably faster at the things you do often — and that library is portable to any AI product that supports the skill standard (and they all will).

Next up

You now have the five foundational building blocks: setup, memory, connectors, projects, and skills. The rest of the course shows these combined on real workflows — Word docs, PowerPoints, Excel models, scheduled tasks that run without you. Start with the scheduled tasks module when published.