Show, don't tell
On July 21, 2026, Anthropic launched Record a Skill inside Claude Cowork — the desktop version of Claude that works directly against your files and tools. The feature flips the logic of how you teach Claude a workflow. Instead of writing a detailed instruction, you do the task once on screen while narrating what you're doing and why — and Claude builds a reusable Skill out of it.
We should be upfront about something right away: Anthropic hasn't yet published official documentation for the feature — no dedicated help page, no deep-dive article. What we know comes from early users, tech press, and Anthropic's own launch announcement on X. That's worth mentioning because it's part of the story: this is a feature shipping into production before the governance framework around it has been fully written. Interestingly, this happens almost simultaneously with OpenAI's similar feature, "Record and Replay," which shipped in June — two companies independently arriving at "show me" as a better entry point to automation than "describe it to me."
How it actually works
The feature sits under the "+" menu in Claude Cowork, available on Pro, Max, and Team plans, currently desktop-app only. You select "Record a skill," perform the task while screen activity, mouse clicks, keystrokes, and your own narration are captured simultaneously, and Claude translates the recording into a structured Skill — the same underlying format as Anthropic's broader Skills system, just a third way of creating one: not hand-written, not generated from a text description, but demonstrated.
The narration isn't decoration — it's where the rules and exceptions actually get captured. A click shows what happens. Your voice shows why. After the recording, Claude reads back what it understood, and builds the actual Skill file across several parallel sub-tasks: one drafts the procedure itself, one writes reference rules specific to the context, one packages the final result. Before you trust it, it should be tested in a fresh conversation via a dedicated command — the same way you'd test a new hire on something simple before handing them the keys to everything.
Two examples that already exist
Marketing consultant Charlie Hills recorded his own social media routine: pulling a caption and image from a content calendar in Notion, and posting it to Buffer. The resulting Skill now adapts the copy automatically per platform — shorter and sharper on X, longer and more detailed on Facebook — and runs unattended every weekday morning at eight. He notes one practical detail worth carrying over: since the Skill actually drives a real browser, the machine needs to be awake and logged in for it to run.
Learning consultant Philippa Hardman used it for something entirely different: not automation, but analysis. She had an experienced employee perform a task while narrating what they were doing, and got an immediate, structured task analysis out of it — something that normally takes hours to map manually. She recorded both a novice and an expert solving the same task, and used the difference to find concrete gaps in the training material. It's a good example that Record a Skill isn't just an automation tool — it's also a tool for making tacit knowledge visible.
How other professions can use the same pattern
The pattern is the same across every example below: do a task once with full narration of the rules, and let the Skill repeat the structure next time. This is our own assessment of realistic use cases based on how the feature actually works — not something Anthropic itself has demonstrated.
HR and recruitment: Screen one batch of CVs against a fixed rubric while narrating out loud why each candidate passes or fails, and finish by writing the rejection letter. The Skill can then run the same assessment and wording on the next batch.
Legal and compliance: Go through one contract clause by clause against a checklist, narrating why a phrase is flagged as non-standard. The Skill repeats the same checklist review on the next contract.
Finance and accounting: Perform one monthly reconciliation — pull the bank statement, match line items against the ledger, explain how discrepancies are categorized. The Skill repeats the matching and flagging routine every month.
IT and operations: Document one incident start to finish — pull logs, fill out the template, explain how severity is assessed. The Skill builds the same report structure the next time something breaks.
Sales: Log one customer call in the CRM right after the conversation — outcome, new deal stage, follow-up email from a template. The Skill repeats the update after future calls.
Customer support: Categorize one ticket per the support playbook and write the response in the right tone. The Skill does the same triage and drafting on new tickets.
The common thread isn't the industry. It's that the task repeats, and the rules for doing it correctly can be said out loud while you do it.
The gap nobody has closed yet
Here's the part we can't not mention, given what we do.
A screen recording captures everything visible — not just the steps Claude needs. Email threads, customer data, dashboards, password managers, Slack notifications, pricing information. Anthropic's only guidance so far on avoiding exposure of sensitive information during a recording is manual — there's no automatic redaction or masking of what gets captured. There's also no published documentation on how long recordings are retained, or exactly what approval process a Skill goes through before it's shared across a team.
The genuinely uncomfortable point is the access gap: Pro and Max plans have essentially no organizational security controls. The Team plan has basic admin controls with permissive defaults. Only at the Enterprise tier do you get SSO, access governance, and real audit capability. That means the feature that captures your most sensitive process knowledge is most freely available exactly where your governance is weakest — on a single employee's personal Pro account, with no one in the organization aware it's happening.
This isn't an argument against adopting the feature. It's an argument for someone deciding how it should be used, before someone else accidentally decides it for you.
What we recommend before you adopt it
Set one simple rule before anyone starts experimenting: never record with real customer data, real personal data, or logged-in systems visible on screen. Use anonymized sample data when building a Skill, the same way you would in a test environment.
Decide who's allowed to publish a Skill to the rest of the team, not just who's allowed to create one for personal use. One person's private shortcut and an organization's standard procedure are two very different things, and should be treated as such.
And consider whether you need Enterprise tier before rolling this out broadly, precisely because that's where the real access controls exist today. It's the same judgment call behind how we recommend rolling out Claude Cowork generally: the functionality isn't the problem. The absence of a decision about how to use it, is.
AI doesn't start with technology. It starts with structure.