"Claude can learn a task by watching you do it once," @DylanAsdaleAI wrote on October 20. "In Cowork on Claude for Mac: + → Record a skill. Do the task, talk through it, click Done. Claude proposes a reusable skill you review and save." Posts like it filled X all week in English and Chinese, even though Anthropic first announced the feature in July. Owners are discovering it now because it finally sounds like something they can use without writing a prompt.
That reaction is fair. Record a Skill removes the hardest first step for a non-technical team: describing a process in writing. But a recording only captures what you did on a good day. The question that decides whether a skill saves time or creates cleanup is whether it knows what a bad day looks like. One of the clearest answers this month came from a financial controller who has built about 50 Claude skills for his month-end close.
TL;DR
- Record a Skill lets Claude watch you do a desktop task once, with narration, and turn it into a reusable skill you review and save.
- It is in the Claude desktop app's Cowork mode, under the + menu, on Pro, Max and Team plans.
- Start with a task you already do the same way every week, and make the first version match today's output exactly.
- Teach it what wrong looks like, add checks that can actually fail, and keep a person approving anything that posts. AI drafts, your team approves.
What Record a Skill actually does
According to Anthropic's announcement, as reported by Search Engine Journal and Digit, you open the Claude desktop app, go into Cowork, choose Record a skill from the + menu, perform the task while talking through it, and stop. Claude uses the screen activity and your narration to draft a skill: a saved set of instructions it can run again later. You review the draft before saving it.
A few facts worth knowing before you plan around it:
- Plans. Pro, Max and Team. It is not on the free plan.
- Where. The desktop app, inside Cowork. It is not a browser extension.
- It is not unique to Claude. Search Engine Journal notes OpenAI's Codex has offered a similar Record and Replay on Mac since June.
- The output is a document. A skill is written instructions, which means you can read it, edit it and correct it. That matters more than the recording.
Why the recording is the easy part
In an X Article titled "Teaching Claude What Wrong Looks Like," @BooksAndBotsAI, a financial controller with about 21 years in accounting, describes building around 50 Claude skills across balance sheet reconciliations, accruals, the CEO month-end deck, budget versus actual and a 13-week cash forecast. His summary of where the effort goes: "The hard part wasn't the AI, it was teaching it what wrong looks like in our books."
He breaks every skill into the same parts: where the data comes from, what the output must look like, what has to tie and to what, what to do when something does not tie, and what it is never allowed to do. He calls the last one the most important.
A screen recording captures the first two well. It shows which report you opened and where the file landed. It rarely captures the other three, because on the day you record, everything tied. That gap is where owners get burned: a skill that runs perfectly on clean data and confidently produces nonsense on the one week a vendor double-billed.
Pick the right first task
Not every repeated task is a good first skill. Use this to sort candidates:
| Task trait | Good first skill | Wait |
|---|---|---|
| Output format | Same report or file every time | Different every time |
| Data source | One or two systems you can read | Data scattered across inboxes and phones |
| How you know it is right | It ties to a number you trust | "It looks about right" |
| What it touches | Read-only, produces a draft | Posts, sends, pays or deletes |
| Frequency | Weekly or monthly | Twice a year |
For a 20-person service business, good first candidates are the weekly job-costing summary, the Monday pipeline report, the month-end deposits reconciliation, or the "jobs completed but not invoiced" list. Poor first candidates are anything that emails customers or moves money.
How to record and harden a skill
Here is the sequence we recommend, adapted from the controller's two phases:
- Choose a report you already produce. Not a new one. You need a known-good answer to compare against.
- Clean the screen before you record. Close client files, banking tabs and private messages you do not want captured. Use a test record if one exists.
- Record it once, narrating the why. Say out loud why you skip certain rows, which column you check first, and what would make you stop.
- Read the drafted skill line by line. Fix anything Claude inferred wrong. This is a document now, so edit it like one.
- Phase one: exact match. Run it against last month and compare to the version your team made by hand. The controller's rule was "boring on purpose": if it cannot match what a good employee produces, nothing else matters.
- Phase two: teach it what wrong looks like. Every time it makes a mistake, add a rule with the reason. His examples include old excluded balances creeping back in and a ship date that passed but turned out fine.
- Add checks that can fail. Tie totals to a source number. Then prove the check fires on a month where you know something was off.
- Keep a person on anything that posts. Reports can run on their own once trusted. Entries, payments and customer messages wait for approval.
The check that cannot fail
The most useful line in the controller's piece is this: "A check that cannot fail is worse than no check, because it launders a bad forecast as a PASS." His 13-week cash forecast had a sanity panel where nine of ten checks passed on a forecast that was clearly too optimistic. Half the checks compared a number to something calculated from that same number, so they could never fail.
His fix was a separate audit skill that shares no logic with the forecast, and a rule that every check must prove on historical data that it can both fire and pass. You do not need his spreadsheet sophistication to borrow the idea. For any skill, ask: if the input were wrong, which line of the output would show it? If the answer is none, the skill is not done.
Preparer and approver
He also draws a clean line between two lanes. Reporting, which is read-only, is where he lets Claude run largely on its own once trusted. Anything that posts to the general ledger is "preview first": the skill builds the entry and backup, runs duplicate and balance checks, and shows him the preview. Sensitive items such as payroll never post without his approval. His shorthand: "Preparer and approver can't be the same, whether the preparer is a person or a bot."
That is the same rule we build every client workflow around: AI drafts, your team approves. Record a Skill makes the drafting cheaper to set up. It does not change who should sign.
He reports the setup saves him roughly 20 to 30 hours a month, and that most of his time has moved from building files to reviewing what they say. That is the realistic win for a small business too. The hours come back, and they come back as review time, not as no time.
FAQ
What is Claude's Record a Skill feature?
It lets you record your screen while you do a task and talk through it. Claude turns the recording into a written, reusable skill that you review and save, then run again later.
Which Claude plans include Record a Skill?
Anthropic made it available on the Pro, Max and Team plans, in the Claude desktop app's Cowork mode. It is not available on the free plan.
Is it safe to record screens with client data on them?
Treat the recording like any screen share. Close client files, banking tabs and private messages you do not need, and use test records where you can. Then read the drafted skill to confirm nothing sensitive was written into it.
What tasks should a small business teach Claude first?
Read-only, repeatable reports with a known-good answer, such as a weekly pipeline summary, a deposits reconciliation, or a list of completed jobs not yet invoiced. Avoid anything that sends, pays or posts until the skill has earned trust.
How do I know a skill is producing correct output?
Run it against a month you already finished by hand and compare. Then add checks that tie to a trusted number and confirm those checks fail on a month where you know something was wrong.
Does OpenAI have a similar feature?
Yes. Search Engine Journal reports that OpenAI's Codex has offered a Record and Replay feature on Mac since June 2026.
Should a skill ever post entries or send messages on its own?
Only routine, low-risk items, and only after its checks pass. Anything sensitive, like payroll, payments or customer messages, should wait for a named person to approve it.