Do it once, on camera: from screen recording to a skill your whole team can run

In July, Anthropic added something quietly useful to Cowork, the agent workspace in the Claude desktop app: you can now record your screen while you do a task, and Claude turns the recording into a skill — a written, reusable set of instructions it can follow the next time that task comes up. You demonstrate; it documents.

That matters because the hard part of skills was never the format. A skill is just a folder with a `SKILL.md` in it. The hard part is that the person who knows the workflow best is usually the person least likely to sit down and write it out. Recording flips that: the fastest way to author a skill is now to simply do the job once, out loud.

What actually happens when you record

The mechanics, per Anthropic’s documentation: in Cowork, click the + button (or Customize → Skills) and choose Record your screen. Do the task the way you normally would — recordings run up to about ten minutes, with a countdown warning near the end. When you stop, Claude analyses the recording and proposes either a new skill or an update to one you already have. You review the draft and click Save or Dismiss.

• Narration counts. If you leave the mic on and explain what you are doing, that context ends up in the skill.
• The video and audio are discarded after processing; only screenshots are retained in the Cowork task.
• It is currently available in Cowork on Claude for Mac, on Pro, Max and Team plans — not yet on Windows or in regular chat.
• macOS will ask for Accessibility and Screen Recording permissions the first time.

Getting a recording that makes a good skill

A skill is only as good as the demonstration behind it, and a few habits make the difference between a draft you save and a draft you dismiss:

• Say why, not just what. “I check the totals tab first, because the export is sometimes a day behind” teaches a decision rule. Silent clicking teaches a click.
• One task per recording. Ten minutes is enough for one workflow done properly, not three done vaguely. Small, single-purpose skills are the ones that stay trustworthy.
• Close everything sensitive first. Anthropic’s own warning is blunt: don’t type passwords or secrets, and don’t have private conversations on screen while recording.
• Narrate the edge cases you hit. If a record is missing or a value looks off during the demo, handle it out loud — that is exactly the judgment a written skill usually lacks.
• Review the draft like a colleague’s first attempt at documenting your job. Claude will occasionally infer a rule you don’t actually follow. Fix it before saving, not after it misfires.

The catch: the skill stops at your laptop

Here is the limitation nobody mentions in the demo videos. The skill you just recorded is saved to your Claude, on your machine. The teammate who asked you “how do you do that report again?” — the whole reason this workflow was worth capturing — still has nothing. They can re-record their own version, and now your team has two slightly different truths about the same process.

Process knowledge trapped on one laptop is the old problem in a new file format. The recording feature solves authoring; it does not solve distribution.

Cowork speaks MCP, so the library doesn’t have to live on your laptop

The good news is that Cowork supports custom connectors — remote MCP servers you add yourself under Customize → Connectors, with just a name and a URL. That is standard plumbing, and it is exactly how shareskills connects: your team’s skill library becomes a set of tools Claude can call directly from the same app you recorded in.

Once the connector is added and you have approved the browser sign-in, the loop closes in plain conversation:

• You: record the task, review the draft, save the skill.
• You: “Publish that skill to our team library on shareskills.”
• Your teammate, in their own Cowork: “Search shareskills for the month-end report skill and install it.”
• Next month, either of you: your agent checks for updates before running, so everyone executes the current version — not the copy from whenever they first heard about it.

Team roles carry through, too. Anyone can publish their own recorded skills; edits to someone else’s skill become suggestions for a maintainer to approve. The recording is how knowledge gets out of someone’s head — the shared library is how it stays one version of the truth.

One editing pass before you share

An honest caveat: a recorded skill captures your context along with your process. Your folder paths, your client’s name in a spreadsheet tab, the quirks of your own setup. Before publishing to the team, do one pass to strip the specifics and turn them into parameters — “the client’s project folder” rather than a path only you have. This is a two-minute job, and you can ask Claude to do it: “generalise this skill so a teammate could run it on a different client.”

The pattern underneath all of this is old: watch the person who is good at the job, write down what they do, hand the write-up to the team. Recording makes the first two steps nearly free. A shared library makes the third one stick.