Every business eventually notices it’s been onboarding the same employee over and over. You explain how your quotes are formatted. Next week, you explain it again. The week after that, same conversation, slightly worse mood.

That’s what using AI without Skills feels like. A Claude Skill fixes it: a small folder of instructions you write once that Claude picks up on its own whenever the work calls for it. Here’s what goes inside one, how Claude decides to open it, and why it’s the difference between an AI that sounds smart and an AI that sounds like your company.

What a Claude Skill actually is

A Claude Skill is a folder of instructions and supporting files that Claude loads when a task matches what the Skill is for. In its original announcement of the feature, Anthropic described Skills as “custom onboarding materials that let you package expertise” — which is about as honest a description as you’ll get. It’s a binder, not a lecture.

Think about the difference between training a new hire and handing them the binder that lives at their desk. The lecture happens once and evaporates. The binder sits there, gets opened when it’s needed, and says the same thing on Tuesday that it said in March.

A prompt is the lecture. A Skill is the binder. And the important part for a business owner: a Skill is just text files in a folder. Not code to maintain, not an integration that breaks when a vendor ships an update. If you can write a one-page document explaining how your company does a thing, you can write a Skill.

What goes inside the folder

Every Skill is a folder with one required file at its center, called SKILL.md. According to Anthropic’s Agent Skills documentation, that file opens with a short block of metadata (two required fields: name and description), and everything after it is plain instructions written in normal language.

The name field is just the Skill’s ID — lowercase letters, numbers, and hyphens only, capped at 64 characters. The description field does the real work. It’s the only part Claude reads before deciding whether to open the Skill, and it has to say both what the Skill does and when to use it:

• Too vague: description: Helps with quotes.
• Actually useful: description: Applies the company house style to customer quotes and proposals, including pricing structure, standard inclusions and exclusions, and warranty language. Use whenever drafting or reviewing a quote, estimate, or proposal.

Matching happens on meaning rather than exact keywords, so write it the way you’d brief a new hire. Anthropic’s own guide to building Skills calls the description “the most critical component” of the whole file — a claim that lines up with how the matching actually works.

Beyond that one required file, a Skill folder can carry whatever the job needs:

• Additional instruction files: a reference document full of detail that would clutter the main file.
• Executable scripts: Python or JavaScript for steps that should run identically every time.
• Resources: templates, example documents, schemas, or API notes.

All three are optional. A perfectly useful Skill can be one file and nothing else, and most good first Skills are exactly that.

How Claude decides to open it

Claude doesn’t read every Skill you own on every request, which is exactly why this scales. Anthropic’s documentation calls the mechanism progressive disclosure, and it runs in stages: Claude always keeps each Skill’s name and description in context (roughly 100 tokens apiece), opens the full instruction file only when a request matches, and reads any bundled reference files only if the instructions point at them.

It’s a restaurant menu, basically. The server has the headings memorized. They walk back and pull the full recipe card only when someone orders the fish.

So you can keep a dozen Skills installed without slowing anything down. They stack, too — Claude can recognize that a task needs two Skills at once and use both.

Why this matters for a small business

The business case isn’t “AI writes faster.” It’s that a Skill converts tribal knowledge into something that runs automatically. The plumbing company owner who knows exactly how a quote should be worded is currently the only place that knowledge lives. A Skill moves it out of their head and into a file that every draft passes through, whether the owner is in the room or on a roof in August.

It also fixes the consistency problem that quietly kills most AI rollouts. Five people using five different prompts produce five different voices under one logo. Five people sharing one Skill produce one voice, without anyone policing it.

Here are three that map onto work small and mid-sized service businesses actually do every week:

Notice what these have in common: each one is a rule set that already exists somewhere in the business, usually in one person’s memory or a Word doc nobody opens. The Skill isn’t new knowledge — it’s knowledge that finally gets used every time. Anthropic’s build guide even recommends writing success criteria into a Skill’s instructions, so Claude can check its own output against a defined standard. That’s a quality gate you get for the price of writing it down once.

Where Skills get built — and the catch

Skills work across Anthropic’s products, but they don’t travel between them on their own, and that trips people up. Anthropic’s help center lists Skills as available on Free, Pro, Max, Team, and Enterprise plans (code execution has to be switched on first), managed under Customize then Skills, with custom ones uploaded as a ZIP file. Developers working in Claude Code instead keep Skills as plain folders in ~/.claude/skills/ or a repository’s .claude/skills/, where committing the folder shares it with the team.

Anthropic’s documentation is explicit about the catch: custom Skills uploaded to one surface don’t automatically show up on the others. Copying a folder isn’t hard, but it’s the kind of detail that makes a rollout feel broken if nobody warned you.

That gap is starting to close. As of a Claude Code update released September 17, 2026, terminal sessions signed in with your claude.ai account now sync the Skills and plugins you’ve enabled there automatically — a setting you can turn off if you’d rather manage each surface by hand. It’s a narrower fix than full cross-platform sync, but it’s real movement on the exact friction point businesses run into first.

What makes a first Skill worth building

Start with something you’ve already done to death. Anthropic’s build guide offers a test worth stealing: have you done this task at least five times, and will you do it ten more? If yes, it’s a Skill. If no, it’s a prompt, and writing a Skill for it is busywork dressed up as automation.

Keep the first one narrow and boring. One document type, one procedure, one checklist. A Skill that tries to cover your entire operation triggers at the wrong moments and produces mush.

Write it, use it on real work for a week, then fix what annoyed you. That’s the entire loop.

Where this leaves you

A Claude Skill is the point where AI stops being a clever assistant you have to supervise and starts being a system that knows how your company works. Writing the first one isn’t a technical project — it’s a documentation project, and most businesses are closer to ready than they think.

Want help turning your standards into Skills that fire at the right time? That’s what our AI Skills and Integrations work covers, from writing the Skill to wiring it into the systems where the work already lives. Not sure which process to package first? Get in touch through contact and we’ll help you sort it out.

Frequently Asked Questions