Claude Cowork Can Now Learn Your Workflow by Watching Once

TL;DR. On July 21, 2026, Anthropic added Record a Skill to Claude Cowork: record your screen while walking through a task and narrating what you're doing, and Claude turns it into a saved workflow it can run again on its own. Available to Pro, Max, and Team subscribers via the + menu in the Claude desktop app.

The hardest part of getting AI to handle a recurring task isn't the automation — it's translating what you do into instructions detailed enough for an AI to follow. Claude Cowork's new Record a Skill feature skips that step.

What Record a Skill does

Open the Claude desktop app, go to your Cowork workspace, and tap the + menu. Choose Record a Skill. Claude begins capturing your screen, mouse clicks, keystrokes, and voice.

Then you do the task.

Walk through it the same way you always do, and narrate as you go — not a formal script, just what you'd say to a new colleague: "I'm downloading this month's report, renaming it with the project code, then moving it to this client folder." When you're done, stop the recording.

Claude processes everything — the screen activity, the click sequence, and your spoken reasoning — and synthesizes it into a Skill: a named, saved workflow it understands well enough to run on its own the next time that task comes up.

The result: instead of writing and maintaining a detailed instruction document, you show Claude once and it extracts the steps from the demonstration.

Who can use it

Record a Skill is available to Pro, Max, and Team plan subscribers. It lives in the + menu of the Claude desktop app only — not the web app, not the mobile apps, even though Cowork expanded to web and mobile on July 7, 2026.

If your workflow involves local files or computer control (features that also remain desktop-only), you need the desktop app regardless. For cloud-only workflows routed through Google Drive, Slack, Gmail, or web-based connectors, the web Cowork works fully — but creating skills from recordings still requires the desktop.

What it's genuinely good for

The feature works best for recurring, predictable tasks — the "I do this exactly the same way every week" work that's tedious precisely because it never changes.

Good candidates:

• Weekly report formatting: download, rename, reorganize, summarize — same steps every time
• Standard content upload workflows with consistent metadata fields (file name, description, tags, thumbnail)
• Recurring data entry that pulls from one source and populates another in a fixed sequence
• Client onboarding checklists that follow identical steps for every new client
• Email triage or classification following a consistent set of rules

Less suited:

• Tasks that vary meaningfully between runs — different recipients, different formats, source data that changes structure
• Workflows requiring real judgment about content: reading a document and deciding what to do based on what's in it
• Exception handling or branching logic that the recording didn't demonstrate
• One-off tasks: the overhead of recording isn't worth it for something you'll only do once

One caveat worth taking seriously: a recorded Skill is only as reliable as the demonstration. If a UI changes, if the task hits an unexpected state, or if a step has a branch the recording didn't cover, the Skill may fail silently or take an unintended path. Treat new Skills as pilots and check their outputs for the first several runs before trusting them unsupervised.

What to keep off-screen

The most important thing to know before you hit record: the system captures everything visible on your screen during the session.

That includes:

• Passwords. If you log into a service mid-recording, that credential is in the session.
• Client data. Any name, address, financial figure, or personal information visible on screen gets captured.
• Regulated information. HIPAA-covered records, attorney-client privileged material, financial data — if you wouldn't screenshot it, don't let it appear during a recording session.

Anthropic has not released detailed privacy or data retention documentation specific to skill recordings as of the July 21, 2026 launch. Until that documentation exists, the safest approach is to record only workflows that don't surface sensitive data, and to do a first run on dummy data before recording a real workflow that touches anything you'd consider confidential.

What this changes

Most AI workflow automation runs into a creation problem: the person who knows a workflow well enough to automate it usually isn't the person who can write the detailed instructions an AI needs. A senior analyst knows exactly how to prepare the weekly reconciliation. Writing it down precisely enough for an AI to follow takes longer than just doing it.

Record a Skill closes that gap. The expert performs the task — once, narrated — and Claude derives the automation from that demonstration rather than from written documentation.

That doesn't make it plug-and-play. Testing and validation still fall on the person deploying the Skill. But the creation step — explaining to an AI exactly how you do something — gets significantly easier when showing it once is enough.

For Teams accounts, Skills can be shared across the team: one person records the workflow, everyone can invoke the Skill. That's the practical multiplier for operations teams with standardized processes.

Sources

• Claude Cowork learns new skills through screen recordings and voice-over explanations — The Decoder (July 21, 2026)
• Anthropic's Claude Cowork Lets You Teach AI by Recording Your Screen — AlphaSignal (July 21, 2026, citing official @claudeai announcement)
• Now You Can Teach a Skill to Claude by Just Recording Your Screen — CyberSecurity News (July 21, 2026)