What Is a Claude Code Skill? (And Why It Changes How AI Builds)

Here's the fastest way to understand a Claude Code skill.

You know Dory, from Finding Nemo — friendly, capable, and unable to remember anything for more than a few seconds? That's an AI assistant without a skill. Brilliant, eager, ready to help — and starting from zero every single time. "Hi! What are we building?" Every. Single. Session.

A skill is the memory you hand it.

The actual definition

A Claude Code skill is a packaged set of instructions, patterns, and reference material you give the AI up front — so it already knows how to do a specific kind of task before you ask. Your rules. Your conventions. The fixes you've already figured out. The stuff that, without it, you'd have to re-explain on every new chat.

A model on its own is general knowledge. A skill is your knowledge, made available to the model exactly when it's working on that kind of problem.

Why a skill beats a better prompt

Most people try to fix bad AI output by writing a better prompt. And prompting matters. But there's a ceiling to it, because a prompt is a one-time instruction — it doesn't persist, and it doesn't carry the dozens of small, hard-won rules that separate "looks right" from "actually works."

A skill persists. It's not "remember to do X this time" — it's "you always know X." That difference compounds. Instead of spending your prompts teaching the AI the basics of your domain, you spend them on the actual work, because the basics are already loaded.

Think of it as the difference between an assistant who is smart, and an assistant who is smart and has done this exact job a hundred times. The second one doesn't need to be told that the manifest needs every required element, or that this particular host can't compile a build step, or that the error message lying to you actually means something else. They just know. A skill is how you give the AI that "they just know."

What goes into a skill

A good skill usually bundles a few things:

• Rules — the non-negotiables for doing the task correctly.
• Patterns — the right way to structure the common cases.
• Reference material — the details the model can pull from when it needs them.
• Examples — working instances it can learn the shape from.
• Fixes — the specific errors that come up, and how to resolve them.

The magic isn't any single piece. It's that all of it is present before you prompt, so the model's first attempt is already informed instead of guessing.

What changes when you use one

The clearest way I've seen this is to run the same prompt twice — once without a skill, once with — on a task that has a lot of unwritten rules.

Without the skill, the AI guesses. The guesses are reasonable, but wrong in the specific ways the domain punishes, and you spend the next half hour in a loop: it produces something, it breaks, you paste the error back, it patches one thing, something else breaks.

With the skill, the first attempt already accounts for those rules. The thing works on the first prompt, and you move on to refining instead of debugging.

Same model. Same prompt. The only variable is what it knew before you asked. That's the whole value of a skill in one sentence.

A concrete example

I build custom Tableau extensions with Claude Code. It's a perfect candidate for a skill, because the domain is full of unwritten rules that punish you instantly: an unforgiving manifest, a sharp hosting trap, an API with quirks, and error messages that rarely say what's actually wrong.

Without a skill, building one is a 40-minute fight. With a skill that encodes those rules, it's a first-prompt build. I wrote up that exact side-by-side — same prompt, with and without — here: Same Prompt, Different Claude: building a Tableau KPI card extension.

And if Tableau extensions are your thing, I packaged that whole skill into a kit: the TableauOps Extension Kit.

The takeaway

Skills are how AI assistants stop starting from zero. They turn a brilliant generalist into a brilliant generalist who also knows how you work. The model is the engine; the skill is the memory. Give it one, and you stop teaching it the basics and start building.

— Eric Summers, Tableau Ambassador & founder of TableauOps