Record a Skill is a feature Anthropic added to Claude Cowork on July 21, 2026, letting users convert a single human demonstration — screen recording plus spoken narration — into a skill Claude can execute again. It differs from writing a custom prompt: a prompt is a text description of steps, while a recorded skill is a physical demonstration that the model itself breaks down into an executable procedure. It is also distinct from a plugin — a skill is one repeatable procedure, while a plugin bundles a full role together with a set of tools. The two sit at different levels, a distinction that technical coverage of the launch specifically flagged as easy to conflate.
The reason this feature exists points directly at a long-standing bottleneck in automation tools: getting an AI to precisely replicate a workflow used to require someone repeatedly describing and refining steps in text, and that process itself carried a real barrier — leading many tasks that were worth automating but too fiddly to justify scripting to simply go undone. Anthropic engineer Lydia Hallie described using the feature for exactly this kind of task — workflows she'd always meant to automate but never did, often not for lack of time but because no connector existed for that particular job. Record a Skill replaces 'describing steps' with 'demonstrating steps,' skipping the prompt-writing stage entirely and targeting this long-standing barrier directly.
The workflow runs in two parts. On the user side, open the '+' menu in the Cowork section of the Claude desktop app, select 'Record a skill' to start recording, then perform the task the normal way — clicking through menus, filling fields, moving between applications — while narrating each step out loud. Anthropic's own demo uses expense filing, the kind of task that recurs monthly or quarterly and stays unscripted precisely because it feels too fiddly to justify the effort. On the model side, Claude simultaneously captures screen activity, mouse clicks, keyboard input, and voice narration. Once recording ends, it processes the whole demonstration into a structured skill definition and saves it, named something like 'Saved as /file-expenses,' after which it can be invoked directly like a built-in command — no need to re-record or re-describe it each time.
If your work includes a task that recurs monthly but has always felt too fiddly to justify scripting, this feature targets exactly that scenario and is worth evaluating. Before rolling it out, though, confirm three things. First, plan tier — it's currently limited to Pro, Max, and Team, with no access on Free. Second, operating system — macOS only for now, with Windows still in planning. Third, the privacy boundary — recording captures screen, clicks, keyboard, and voice together, and as of launch Anthropic had not published detailed retention policy for skill recordings, so anything touching passwords, client PII, or regulated data should be deliberately kept off-screen; don't assume automatic redaction. Reliability is also still early-stage: whether a workflow inferred from one demonstration holds up reliably as screen layouts or data content change is currently something users are mostly validating themselves.
On July 21, 2026, Anthropic added Record a Skill to Claude Cowork, lowering the barrier for 'teaching Claude to do something' from writing a prompt to simply demonstrating the action. Users start a recording, perform the task the way they normally would, and narrate each step out loud. Claude then breaks the recording down into a reusable skill, saves it to the skill library, and the next time the same task comes up, it can be called directly instead of walking Claude through it again.
The entry point sits in the '+' menu of the Cowork section in the Claude desktop app, the same menu used for attaching files, browsing the skill library, and managing connectors. Once recording starts, Claude captures screen activity, mouse clicks, keyboard input, and voice narration together. Anthropic's own demo on X uses filing an expense report — the kind of task that recurs monthly or quarterly but rarely gets scripted because writing automation for something used once a month never feels worth the effort, which is exactly the scenario this feature targets. Once the recording finishes, the system saves the result as a skill definition, named something like 'Saved as /file-expenses,' and from then on it can be invoked like a built-in command. Once a skill lands in the library, calling it again in a brand-new session the next time the same task comes up requires no re-demonstration and no re-description — that's the session continuity the skill library provides, carrying the effect of a single demonstration forward into every future call.
Anthropic engineer Lydia Hallie described using the feature for workflows she'd always meant to automate but never did — often not because she lacked the time, but because no MCP Server existed for that particular task — historically, getting Claude to operate an external service required an already-written MCP Server acting as the bridge, and anything without one had to stay manual. Record a Skill sidesteps that requirement entirely: no one has to write the connecting interface first, because Claude learns directly from the demonstration. That points at what the feature actually solves: getting an AI agent to precisely replicate a workflow used to require repeatedly writing and refining a prompt by hand. Now the demonstration happens once, and the model handles breaking it into steps on its own.
A Claude 'skill' is fundamentally a folder containing instructions, scripts, and resources that the model can read in real time to strengthen its performance on a specific task. This connects to a technical distinction reporting on the launch has flagged directly: Record a Skill produces a skill — a single repeatable procedure — not a 'plugin,' which bundles a full role together with a set of tools. The two sit at different levels and shouldn't be conflated.
Worth noting: when a skill involves multiple steps, Cowork's underlying architecture may split the work across several sub-agents that each handle a segment before the results are combined — that's an existing part of how Cowork executes tasks, and Record a Skill only changes how the skill definition gets created, not how it runs underneath. Cowork itself launched as a research preview in January 2026 and has been described as 'Claude Code for people who don't write code.' It expanded to mobile and web in early July, defaulting to cloud execution. Cowork already shipped with a set of built-in skills covering slides, spreadsheets, Word documents, and PDFs, but the barrier to building custom skills remained high; Record a Skill opens a path for non-technical users to build their own, pushing Cowork from a tool that executes existing skills toward one that can extend its own capabilities. Set against the wider category, this is a new on-ramp into workflow automation — a field that has historically relied on writing scripts or configuring trigger conditions, and that this feature replaces with 'show it once.'
The feature is currently available on Pro, Max, and Team plans; Free-tier users cannot access it. It only runs on macOS — a Windows version is still being planned — which limits reach in the near term. Enterprises evaluating adoption should pay particular attention to the privacy boundary: recording captures screen, clicks, keyboard, and voice simultaneously, and as of launch Anthropic had not published detailed retention or privacy policy specific to skill recordings. Anything involving passwords, client PII, or regulated data should be deliberately kept off-screen during recording — don't assume the system will redact it automatically.
Reliability is also still in an early validation phase. Whether a workflow inferred from a single demonstration holds up reliably in real-world conditions where screen layouts and data content vary is currently something users are still testing in the field; Anthropic has not published large-scale stability figures.
If your job includes the kind of task that repeats every month but is too fiddly for anyone to bother scripting — expense reconciliation, a recurring weekly report in a fixed format, month-end account matching — that's exactly the target scenario for Record a Skill. The usual reason people never automated it was that the cost of writing a script once didn't justify a task used only once a month. That cost has now dropped to 'do it once, narrate it once,' which genuinely lowers the barrier. Before rolling this out, though, confirm your plan tier (Pro, Max, or Team), your operating system (macOS only for now), and — before recording any workflow touching sensitive data — decide in advance what stays off-screen. Don't wait for an incident to write the privacy rule after the fact.