All posts

Cowork Series · Apr 3, 2026

Skills in Claude Cowork

Part 2 of the Cowork series: what a skill is, what it consists of, how to build your first one, and how skills chain into whole workflows.

Built by Simon + Claude

TaskAIResearch1Draft2Check facts3FormatFile4NotifyResultThe agent decides which, and in what order.The skills stay small.

TL;DR

  • A skill is a folder with instructions, examples and scripts Claude can run.
  • It turns a generalist into a specialist that knows your methods, templates and tone.
  • Start with one markdown file for the task you repeat most, then chain skills into workflows.

In part 1 we looked at what Cowork is: point Claude at a folder, describe what you want, get finished files back. Impressive, until you say "write a report based on these meeting notes" and get something that looks fine but lacks everything that makes it your report. The angle, the priorities, the format your team recognises.

That is not Claude's fault. It simply does not have the context. That is what skills give you. You capture your methods, templates and rules once, and every later delegation starts from your context instead of generic assumptions.

What is a skill?

Anthropic introduced Agent Skills as an open way to package instructions, scripts and reference material into reusable folders. At its core, a skill is a text file with instructions, frameworks and context that Claude pulls in automatically when it is relevant. Since then the concept has spread across the industry, and other AI tools now support the same idea.

Think of it as the difference between hiring a generalist and a specialist. Without a skill, "write a report" gives you something correct but generic. With a research skill that contains your methodology, source priorities and report template, Claude starts from a completely different place.

What a skill consists of

A skill is really just a folder with up to three kinds of content.

Markdown is the core. A file with instructions, methodology and rules. This is what Claude reads to understand how the task should be solved.

Scripts are where it gets interesting. Small programmes Claude can run as part of the task. You do not need to code; Claude writes and runs them for you. A system prompt can only say "do this". A script can actually do it. The markdown file is the brain, the script is the hands.

Files are reference material. Examples of good output, templates, brand guides, tone of voice documents.

Markdown, brainScripts, handsFiles, referencesOne skill

Most skills start as a single markdown file. Add scripts and files when you discover Claude needs them.

Skills across Claude, Claude Code and Cowork

Skills work across the Claude products, with slightly different journeys. In claude.ai you manage them under Customize and Skills, where Anthropic's built-in skills for Excel, PowerPoint, Word and PDF also live. Cowork supports both standalone skills and plugins, and has a built-in skill-creator that builds skills straight from the chat. In Claude Code, skills are files in the .claude/skills folder, and the whole machine is available, which makes advanced skills essentially unbounded.

The format is the same everywhere: a folder with a SKILL.md file, optional scripts and reference material. Build it once, use it where you need it.

Skill workflows

A single skill does one task well. But because the output of one skill can feed into the next, you can build whole production lines. First skill researches the market. Second turns the research into positioning. Third uses the positioning to write a landing page. Fourth breaks the message into an email series.

01Research skill02Positioning skill03Landing page skill04Email skillOne in, four deliverables out

Each skill has its own methodology and its own quality bar. You start with one instruction and end up with four finished deliverables. That is not prompt writing anymore. That is systems building.

From prompt to system

Most people practice light prompt engineering: phrase the request a little better, add some context, get slightly better output, and start over next time. Skills flip this. Instead of getting better at asking, you build systems that deliver. The difference is like explaining a task to a new temp every week versus having a colleague who already knows how you work.

The first time you build a skill it can feel like overkill. The tenth time you use it, and it delivers consistent quality without you thinking about it, you understand the point. And skills compound: every time you spot a gap you can say "update the skill based on this conversation" and get a better version on the spot.

Build your first skill

  1. Ask Claude to interview you, five to ten questions, and suggest skills that fit what you work on and what takes the most time.
  2. Pick one suggestion and ask Claude to use the skill-creator. Describe how you actually want the task done.
  3. Test it on real work, give feedback, and let Claude update it.

The first version does not need to be perfect. Skills improve through use, not planning.

Tips for the first days

  • Start with one skill, not fifty. Pick the task you do most often.
  • Iterate with feedback. Three rounds of feedback beat three hours of planning.
  • Use the skill-creator. Say what you want and let Claude build.
  • Check the output. Skills make Claude better, not infallible.

We think skills are one of the biggest unlocks available right now, especially inside companies. You can start with a simple skill, use it immediately, iterate while you work, and then share it so the whole team gets the same standard. You do not need to code. You describe how you work, and Claude does the rest.

Skills make Claude smarter about what it can already see. The most important context usually lives outside: in your email, calendar, CRM and knowledge base. That is what connectors solve, and that is part 3.

First published in Norwegian on LinkedIn.