There are two ways to write a bad prompt for an automation. You can give Claude three sentences and quietly expect it to fill in twenty steps you never mentioned, or you can hand it nine hundred words of edge cases and exceptions that bury the actual job. I have done both, usually on the same agent, usually in the same week.
Claude's Record a Skill feature is aimed straight at that problem. Instead of describing a task, you record yourself doing it, talk through it while you work, and Claude builds the skill from what it watched. Here is how it works, who can use it, what Claude keeps afterward, and which jobs are actually worth recording.
Key takeaways
-
Record a Skill turns a screen recording plus your narration into a Claude skill, so you demonstrate a workflow instead of writing instructions for it.
-
It lives in Cowork in Claude for Mac, on Pro, Max, and Team plans. Anthropic's documentation says it is not in chat, not on Windows, and not on Free or Enterprise plans.
-
A recording can run for about ten minutes, and the capture bar counts the steps it has picked up as you work.
-
Claude does not add a skill to your account on its own. It reviews the recording in a Cowork task and proposes a skill, which you either save or dismiss.
-
If the recording overlaps a skill you already have, Claude proposes an update to that skill instead of a new one.
-
The video and audio are not kept. A set of screenshots does stay inside the Cowork task automatically, and deleting the task removes them.
-
Narration is the half that matters most, because your voice carries the reasoning your clicks do not.
-
A saved skill carries no recurring bill beyond the Claude subscription you already pay for, which changes the math on what is worth automating at all.
What is Record a Skill in Claude?
Record a Skill is a way to create a Claude skill by demonstrating a task rather than describing it. You record your screen, your clicks, your typing, and your voice while you do a piece of work you already know how to do. Claude watches the recording, writes a skill from it, and shows you the result for approval before anything gets saved.
A skill, for anyone who has not built one, is a set of instructions Claude loads when it recognizes a matching task. It is how you teach Claude your process for something specific: how your product listings get built, how your weekly report gets assembled, which fields matter and which ones you always skip. Historically, that meant a blank Markdown file and a lot of typing.
The recording path replaces the typing with a demonstration, and the payoff there is bigger than the time you save.
How do you record a skill in Claude Cowork?
Open Cowork in Claude for Mac, start a recording from the "+" menu, do the task while narrating, then click Done and review what Claude proposes. Anthropic's help center article on creating custom skills has the authoritative version. Here is the sequence with the parts that trip people up.
Before you record, update to the latest version of Claude for Mac. The first time you try this, macOS asks for two permissions: Accessibility, so Claude can track your mouse and keyboard, and Screen recording, so it can see your screen. macOS may make you restart Claude afterward. Close anything you do not want captured, because everything on your screen goes into the recording along with everything you say. Do not type a password and do not leave a private conversation open in another window.
There are two ways in, and the second one is easy to miss. The first is the "+" button in the Cowork composer, where you select "Record a skill." The second is Customize > Skills, where you click "Add" and choose "Record your screen." Same feature, two doors.
Click "Start recording." The microphone is on by default, and there is a control to mute it or switch inputs if you need to. Then do the task the way you normally do it. A capture bar sits on screen showing that recording is live and counting the steps it has captured, which is a useful sanity check that it is following along.
Click "Done" when the job is finished, or "Discard" to throw the whole thing away. A recording can run for about ten minutes, which is a ceiling rather than a target. When roughly one minute is left, a countdown appears in the capture bar, and at zero the recording stops and sends everything captured so far to Claude, exactly as if you had clicked Done. Plan the demo to fit inside that window.
What happens after you stop recording?
Claude opens a Cowork task, reviews the recording, and proposes a skill for you to approve. No skill is added to your account until you say so.
You get one of two cards. A brand new skill shows up marked NEW, with Save and Dismiss buttons. If the recording overlaps something you already have, Claude proposes an update to that existing skill instead, and the card tells you which skill it is based on, with Update and Dismiss. Either way, you can expand Content on the card and read the actual skill before deciding.
Read it. That approval step is where the real value shows up. It is the first time you get to see your own workflow written down by something that was paying closer attention than you were. Skills you keep land in Customize > Skills and behave like any other skill from there.
Why is showing better than prompting?
Demonstrating a task beats describing it because most of your process is knowledge you stopped noticing years ago. The bottleneck in building an automation has always been recall.
This is where the feature earns its keep, and it splits neatly across the two ways people get prompts wrong.
The under-engineered prompt
Under-engineering looks like a sentence and a half. "Pull the weekly numbers and put them in the sheet." In your head, that request is complete, because you have done it two hundred times and every decision inside it has gone automatic. Which tab. Which date range. What you do when a row is blank versus when it is zero. The one vendor whose numbers always land a day late, so you wait.
None of that ever gets typed, because none of it feels like a step. It feels like knowing your job. A recording captures it anyway, because you cannot complete the task without doing those things, and the camera does not care whether you thought they were worth mentioning.
The over-engineered prompt
Over-engineering is the louder failure. You sit down to write the perfect instruction set, and because you are being thorough, you start writing rules for situations that have never happened. Three paragraphs on a data format you saw once. A contingency plan for a vendor you stopped using last year. By the time you finish, you have written a document nobody would read, and Claude has to weigh every one of those rules against the two that actually matter.
I feel that pull constantly. We have an agent that wakes up nine times a day, every business day, and reads what the team said in Slack. It runs on a table mapping every kind of incoming message to two decisions, and writing that table was most of the build. That one earned the effort, because deciding when an agent is allowed to speak or act genuinely is the design problem, and we could only write the table after watching a month of real requests come in. For a task that is twelve clicks in a fixed order, that same instinct produces a wall of text where a two-minute video would have done it.
Recording puts a natural ceiling on both. You cannot under-specify, because you have to actually complete the work. You cannot over-specify, because the edge cases you love writing about will not come up during the demo.
Why does narration matter so much?
Narration matters because the screen only shows what you clicked, while your voice explains why you clicked it.
Anthropic makes this point in the documentation, and it is the single tip I would put above the rest. Talk the entire time. Say why you are skipping a field. Say how you chose between two options that looked identical. Say what you would have done if the number had been different. A silent recording gives Claude a sequence of actions, which is the cheap half of your expertise. The narration gives it the judgment underneath, which is the part that is hard to hire for and even harder to write down.
Treat it like training a new teammate who is standing behind your chair, because functionally that is what is happening.
Who can actually use Record a Skill?
Record a Skill is available to Pro, Max, and Team subscribers, inside Cowork, in Claude for Mac. Anthropic's documentation currently states that it is not available in chat, not on Windows, and not on Free or Enterprise plans.
That Windows line deserves more attention than it gets. Cowork itself runs fine on Windows, so plenty of people are going to open the desktop app, go looking for "Record a skill" in the composer, and find nothing there. The feature is missing, not the user. Availability on any of this can change, so the reasonable move if you are on a Mac and on a qualifying plan and still do not see it is to update the app and check the "+" button again before assuming something is broken.
Worth saying plainly rather than writing around: everything above is the documented behavior of the feature, not a case study. My own skills were all written by hand, and the ones I would most want to record are exactly the click-heavy jobs I keep failing to describe properly.
What does Claude keep from your recording?
The video and the audio are not retained. What does stick around is a set of screenshots from the session, stored automatically inside the Cowork task that built the skill.
You can look at them by expanding the "Recorded demonstration" step in that task. Because they live in the task, deleting the task removes them from your task history immediately, and Anthropic's stated retention policy purges deleted tasks from backend storage within thirty days. That is a clean model, and it is worth knowing before you record, mostly because it tells you exactly where the sensitive frames would be if you accidentally captured one. The same rule applies as always: assume everything visible during those ten minutes was captured, and set up your screen accordingly.
Which tasks are worth recording, and which are not?
Record the repetitive, click-heavy work that runs on a stable interface, and keep writing the instructions by hand for anything that turns on judgment.
The good candidates share a shape. The task repeats often. The interface is stable. The steps are the same every time, and the variation lives in the inputs rather than in the decisions. Anything click-heavy qualifies, especially the workflows carrying a dozen small habits you would never think to document. Product listing builds, report assembly, a recurring audit pass, moving data between two systems that refuse to talk to each other.
The bad candidates are the judgment calls. Anything where the real work is weighing tradeoffs, and the clicking is a rounding error, is still better served by writing the instructions carefully. A recording of you thinking is a recording of you staring at a screen.
The other rule I would hold onto: run the skill once on a fresh, real example before you point anything important at it. A recorded skill is a first draft of your own judgment, and one live test tells you which parts survived the translation.
What does it cost to run a recorded skill?
A recorded skill has no recurring cost beyond the Claude subscription you already pay for.
This is the part I would sit with longest. When every automation carries its own monthly invoice, you only automate problems big enough to justify a line item, and the small leaks never get covered because nobody wants to add another bill for them. A skill you recorded in four minutes has no vendor, no seat count, and no renewal. It lowers the bar for what is worth automating, which changes what gets automated at all.
There is one real meter running. Anthropic notes that Cowork tasks consume more of your usage allocation than regular chat does, so a skill that fires constantly will show up in your limits even though it never shows up on an invoice.
Free is not free, and I would rather say that plainly. You pay in the time it takes to do a clean demo, you pay again when the interface changes and the skill starts producing something slightly wrong, and you pay a third time in the attention it takes to notice. The bill is real, and it gets paid in hours.
Frequently asked questions
Where is Record a Skill in the Claude app?
Two places. Click the "+" button in the Cowork composer and select "Record a skill," or go to Customize > Skills, click "Add," and choose "Record your screen."
Is Record a Skill available on Windows?
No. Anthropic's documentation states that recording a skill is available in Cowork in Claude for Mac and is not available on Windows, in chat, or on Free and Enterprise plans. Cowork itself does run on Windows, which is why the gap surprises people.
How long can a Claude skill recording be?
About ten minutes. A countdown appears in the capture bar when roughly a minute is left, and when it reaches zero, the recording stops and sends everything captured up to that point, the same as clicking Done.
Does Claude save the recorded skill automatically?
No skill is added without your approval. Claude reviews the recording in a Cowork task and proposes a skill, which you can read before choosing Save or Dismiss. If it overlaps an existing skill, Claude proposes an update to that skill instead.
Does Anthropic keep the video of my screen?
No. The video and audio from the recording are not retained. A set of screenshots is stored inside the Cowork task that created the skill. Deleting that task removes them from your history immediately, and Anthropic's retention policy purges deleted tasks from backend storage within thirty days.
Do recorded skills cost extra to run?
There is no recurring fee beyond the Claude subscription you already pay for. Cowork tasks do consume more of your usage allocation than regular chat, and the rest of the cost is your time: recording a clean demonstration, then fixing the skill when the interface it was built on changes.
What kind of task should I record instead of writing out?
Repetitive, click-heavy work on a stable interface, where the steps are identical every time. Judgment calls are still better written by hand, because a recording of you deciding something captures very little.
Final thoughts
Most people are terrible at describing their own work, and that has been the real ceiling on building automations. The tooling caught up a while ago. The documentation never did, because in most companies the documentation lives in somebody's head and comes out one forgotten step at a time.
Every skill I have written by hand has the same flaw somewhere in it. There is a step I do without thinking, so I did not write it down, so the agent skipped it, so I found out later when something looked slightly off. Demonstrating the task closes that gap by removing my judgment about what was worth mentioning. A recorded skill also carries no monthly bill, which is what lets the small annoying jobs finally clear the bar for getting automated at all.
Record something small first. A four-minute task you have done a hundred times, narrated out loud like you are teaching it. Then read what Claude writes back, because seeing your own process described accurately by something that just watched you do it is useful whether or not you keep the skill.
Interested in learning more?
If you want the operator side of running agents rather than the feature side, start with How We Built an AI Agent to Catch Discount Code Fraud Every Morning, then How We Built an AI Agent That Asks Permission Before It Acts, and finally How We Let an AI Agent Earn Autonomy One Action at a Time.
For a smaller, self-contained example of automating something nobody else was going to automate, take a look at How I Automated My Monthly Personal Financial Statement. For the other side of the argument, where a tool is genuinely exciting and still not worth installing yet, there is I Love OpenClaw. I'm Not Installing It Yet.
