Every AI chat starts from zero. Explain your brand colors again. Re-paste the SEO checklist again. Remind it, again, that this site has exactly four categories and not five. Anthropic’s answer to that treadmill is Claude Skills — folders of instructions Claude can load on its own when a task calls for them. The pitch is simple: teach it once, not every time. The real question is whether that’s a genuine fix or just a fancier place to paste the same instructions.

What exactly is a Claude Skill?

A Skill is a folder with a SKILL.md file at its center, plus whatever else the task needs — reference docs, checklists, even small scripts. Anthropic introduced them in October 2025 and opened the format up in December 2025, when skill management went org-wide and a partner directory launched with early skills from companies like Box, Canva, Notion, and Rakuten. Box uses one to turn stored files into standardized decks and spreadsheets; Rakuten uses one to process spreadsheets into financial reports. The common thread is a repeatable, well-defined procedure that used to live in someone’s head, or in a doc nobody reliably opened before starting the task.

How does Claude know which skill to use?

Through what Anthropic calls progressive disclosure. Claude keeps a short description of every available skill in view — name and a one-line summary, essentially — and only pulls in the full SKILL.md body when a request actually matches it. A skill about formatting spreadsheets stays dormant during a conversation about travel plans. That matters more than it sounds: the alternative is stuffing every instruction into a system prompt permanently, which costs context space whether you need it that turn or not. A skill costs nothing until the moment it’s relevant.

How is this different from a custom prompt, or an MCP tool?

A prompt you paste in has to be pasted in again next time. A skill persists, and it’s portable — the same folder works the same way in the Claude apps, in Claude Code, and through the API, rather than being locked to one chat window. It’s also a different layer from MCP servers and tools. Tools and MCP connections give Claude new abilities — read this database, call this API. A skill is closer to a process document: it tells Claude how to use the abilities it already has for a specific kind of job, in a specific way, with your specific preferences baked in. The two compose: a skill can absolutely instruct Claude to reach for a particular tool at a particular step.

What does a skill actually look like, in practice?

In Claude Code specifically, a skill is just a directory — .claude/skills/your-skill-name/SKILL.md — with frontmatter describing when it applies and a body full of the actual instructions. It’s close enough to a custom slash command that the two were merged: a file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both produce a working /deploy command. The difference is that a skill can carry supporting files alongside it, and its body only loads into context when it’s actually used — unlike a standing instructions file, which costs tokens on every single turn whether it’s relevant or not.

We used one to write this article — here’s what actually happened

This isn’t a secondhand summary. This site runs on a house-style skill: brand colors, the exact four categories a post is allowed to use, the SEO field lengths, and a fairly detailed image pipeline, all written down once. Asked to write and publish a post, Claude loaded that skill automatically — no one had to re-explain that the cobalt is #013CB0, that headings run in Poppins at 700 weight, or that an SEO description belongs in the 140–155 character range. It was just already there, applied the same way it was the last time and the time before that. The alternative — restating all of it in the chat, hoping nothing got left out — is exactly the treadmill Skills are built to get off of. It’s a small thing until you notice how much of a repeated workflow is actually just repeated instructions.

Is it worth setting up?

If you only ever ask Claude for one-off help, probably not — there’s nothing to repeat. But the moment you notice yourself pasting the same three paragraphs of context into a new chat for the second or third time, that’s the signal. A skill turns a habit you’ve built into something Claude can just have, the same way AI agents more broadly are shifting from answering questions to carrying out multi-step work on their own — skills are part of what makes that work reliable instead of improvised each time.

Before you build one

• Write down the instructions you keep repeating — if you can’t state the procedure plainly, a skill won’t fix that, it’ll just enshrine the confusion.
• Keep the description short and specific, since that’s the only part Claude sees before deciding the skill is relevant.
• Put long reference material in the body, not the description — that’s the whole point of progressive disclosure.
• Test it on a real task before trusting it on an important one, the same as you would any new tool.

Skills aren’t a bigger model or a smarter one. They’re closer to onboarding paperwork — unglamorous, and the reason a new hire doesn’t have to be told the coffee machine rules on day one and every day after. That’s a smaller claim than most AI headlines make, and also a more useful one.