Claude Cowork skills are the same thing your coworker who lives in the terminal has been quietly using for months. A skill is just a folder with a file called SKILL.md inside it, and that one file teaches Claude to do a specific task your way, every time, without you re-explaining it. Cowork is the version of Claude that handles multi-step desktop work for you (sorting files, pulling receipts into a spreadsheet, drafting a branded deck), and it reads those exact same skill folders. So when someone hands you a skill, or you build one yourself, it runs in Cowork, in Claude Code, and across Codex, Cursor, and Gemini CLI too. One format that travels everywhere.
That portability is the part most people miss, so I want to start there.
What a skill actually is, in plain terms
Picture the onboarding doc you'd write for a new hire. "Here's how we format the weekly report. Here's the tone for client emails. Here's the checklist before anything goes out the door." A skill is that doc, written for Claude instead of a person.
The mechanics are boring in a good way. You get a SKILL.md file with two labeled bits at the top, a name and a description, and then the instructions underneath in plain English. Anthropic's Agent Skills overview lays out the whole structure. The description is the one that earns its keep, because it tells Claude both what the skill does and when to use it, so Claude can reach for it on its own when your request matches.
A skill can stay exactly that simple, or it can carry extra files: a template to fill in, a reference sheet, even a small script that does something the same way every time. Those extra files sit in the folder and only get pulled in when the task needs them. You do not have to touch any of that to get value, because a plain-text instruction file is already a complete, working skill.
Claude Cowork skills run the same SKILL.md as Claude Code
This is the claim I'd put money on. The skills you build for Cowork follow the open Agent Skills standard rather than some special Cowork dialect, which means the folder is portable.
Anthropic's own docs say it directly: the same skill runs across claude.ai, Claude Code, and the API without modification. Cowork sits inside Claude's desktop app and reads that same format, so a skill someone wrote in the terminal works for you in Cowork, and a skill you build in Cowork works for the developer down the hall with no conversion and no "export to non-coder mode" step in between.
Why does this matter to you, the non-coder? Because it means you are not locked into one corner of the product. You build the skill once and it travels with you wherever Claude goes.
The built-in office skills you already have
Before you build anything, it helps to know what ships in the box. Anthropic provides pre-built skills for the four formats you actually use: PowerPoint, Excel, Word, and PDF. The overview docs list them as pptx, xlsx, docx, and pdf.
These run quietly in the background. When you ask Cowork to build a deck or drop data into a spreadsheet with working formulas, the relevant office skill handles the formatting without you invoking it or configuring anything. The Cowork product page frames the experience as "hand off a task, get a polished deliverable," and these skills are a big part of why the deliverable comes back looking finished.
So the bar for building your own skill is lower than you'd think. The hard formatting is already handled, and what you add is the part specific to you: your structure, your rules, your style.
How to create a skill in Claude Cowork with no code
Here's the path that does not involve writing a SKILL.md by hand.
Anthropic publishes a skill creator whose entire job is to make other skills. On the Claude skills page it's described plainly: you describe what you want, and Claude generates the folder structure, formats the SKILL.md file, and bundles your resources. You talk, it drafts, and then you review what it produced. That review step is the one that actually matters, because you are the only one who knows whether the instructions match how you really work.
Treat it like briefing an assistant on their first day. Tell it the task in plain language. "I write a Monday status update. It pulls from three sources, it's always five bullets, it never uses jargon, and the subject line follows a set pattern." Be specific about the small rules, because the small rules are exactly the stuff you're tired of repeating. The skill creator turns that brief into a draft skill, you read it, you fix anything it got wrong, and now you've got something reusable.
One honest note on the steps. The exact in-app wording and menu names inside Cowork shift as the product updates, so I'm not going to hand you a click-by-click screenshot that will be stale next month. The reliable mental model is the one above: describe the task in conversation, let Claude draft the skill, review and correct, then save. If you want a deeper walkthrough of the no-code approach in general, we wrote one on how to build a Claude skill with no code.
There's also a second no-code path worth knowing about. Knack builds the same SKILL.md from a short interview, then hands you a finished skill that Cowork runs like any other. It walks you through a structured set of questions instead of an open chat, which some people find easier to think through, and the output is the identical open format. Same destination, different on-ramp.
Whichever path you pick, the result is one portable skill folder. The skill creator drafts it in conversation, while Knack walks you through an interview at getknack.ai and produces the same file. Both leave you with something Cowork, Claude Code, and the rest can read.
How a skill triggers
You have two ways to fire a skill, and Cowork supports both because they're baked into the format.
The first is automatic. Claude reads every skill's description at startup (just the description, which is tiny, not the whole skill). When your request matches one, Claude loads the full instructions and follows them, so if you ask for your Monday update by name or by what it does, the right skill activates on its own. This is why a sharp, specific description is worth more than a clever one. The description is the trigger, and a vague description triggers at the wrong times.
The second way is direct. You can call a skill yourself by its name when you want it on purpose, regardless of how you happen to phrase the request, which is handy for the tasks where you'd rather not have Claude guessing about timing.
This loading-only-when-needed design has a name, progressive disclosure, and it's the reason you can keep a whole pile of skills installed without slowing anything down. Claude only knows each one exists until the moment it becomes relevant. Our primer on what a Claude skill is covers the stages for non-developers.
The one rule that separates good skills from junk
Keep each skill small and single-purpose.
The temptation, once you see how easy this is, is to build one mega-skill that does your whole job. Resist it. A skill that tries to handle status updates and client emails and the quarterly deck and invoice cleanup will trigger at the wrong moments and hand you mush. Claude matches your request against the skill's description, and a description that tries to cover four unrelated tasks ends up describing none of them well.
Build four skills instead. One does the Monday update, one does client emails, and each one carries a tight description so it fires when it should and stays out of the way otherwise. Small skills are also far easier to fix when one starts to drift, because you can read the whole thing in thirty seconds and spot what's wrong. A handful of narrow specialists will beat one overloaded generalist every time.
The first skill you build should be the most boring, repetitive task you do every single week. Build that one, watch Cowork nail it three times in a row, and you'll understand why people end up with a drawer full of them. For the broader context, see what Claude Cowork is. Then go automate the thing you're already sick of doing by hand.