Claude Can Now Watch You Work and Turn It Into a Reusable Skill
Claude's new record-a-skill feature lets you demonstrate a task once so it can repeat it automatically. Here's how it works and where it breaks.
What is Claude’s record-a-skill feature?
Claude’s record-a-skill feature lets you demonstrate a task on screen once, while narrating your reasoning out loud, so Claude can turn that recording into a reusable, repeatable workflow. Instead of writing a prompt or building an automation flow, you just do the task in front of Claude Cowork’s desktop app. Claude watches your screen, listens to your mic, and converts what it sees into a saved “skill” you can trigger again later with a slash command.
TL;DR
- Screen-and-voice recording replaces prompt engineering: you show Claude a task once, explain your decisions out loud, and it packages the process into a skill.
- The feature only works inside Claude Cowork’s desktop app, not the web version, because it needs access to your screen, mic, and local files.
- Two demonstrated use cases stood out: running a QA checklist against a YouTube video (thumbnail, description, chapters, playlist, end screens) and automating an invoice processing workflow that sorts files, extracts fields, and renames documents.
- A 10 megabyte file size limit currently blocks video-heavy workflows, so this is better suited to documents, spreadsheets, and web-based checks for now.
- Skills run through a browser extension that uses Opus 4.5, independent of whichever model you have selected in the chat window.
- You can attach an existing checklist or SOP document during recording instead of narrating every rule, which speeds up setup considerably.
- A related but separate update bumped ChatGPT’s custom instructions limit from 1,500 to over 5,000 characters, making it more practical to paste long personal context files directly into personalization settings.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How do you set up record a skill?
You need the Claude Cowork desktop app, not the browser version. If you only have web access, go to the apps and extensions menu inside Cowork and install the desktop application from there.
Once installed, switch into Cowork mode (this is what gives Claude access to your local files during recording) and look for a plus icon. Under it you’ll find “record a skill.” The first time you use it on a Mac, Claude asks for screen and microphone permissions, since it needs both to observe what you do and hear you explain it.
Before recording, there are a few setup steps worth doing:
- Go to settings and enable code execution and file creation under capabilities. This is required for skills to function.
- If your workflow touches a website, you may need to install and update a companion Chrome extension and grant it permissions.
- Close any tabs or windows you don’t want Claude to see. Once recording starts, it takes in everything on screen as context, and stray tabs can confuse the resulting skill.
After hitting record, you perform the task normally: click through the steps, narrate why you’re making each decision, and stop when you’re done. Claude then asks whether to save this as a new skill or fold it into an existing one, packages the recording, and stores it. You trigger it later with a slash command tied to the skill’s name.
What can you actually automate with this?
Two demonstrated workflows show the range: quality checks and document processing.
QA checklists. If you have a repeatable checklist for reviewing work (a video upload, a document, a piece of content), you can show Claude the checklist itself rather than reciting every rule verbally. In one example, a checklist covered things like whether a thumbnail’s text was readable, whether the video had been added to a playlist, whether chapters were included, and whether placeholder text had been removed from the description. Claude opened the relevant dashboard, checked each item against the actual video, and returned a short report flagging what passed and what was missing, including specifics like “no end screen or cards added” and “no sponsor mentioned.”
Document and invoice workflows. In a second example, a folder structure held three subfolders: incoming invoices, a needs-review folder for anything that failed quality checks, and a processed folder for invoices that were clean and ready to log. The recorded process opened each invoice, checked for required fields (document type, vendor, document number, date, due date, currency, subtotal, tax, total), applied a consistent file naming convention based on date and vendor, and moved each file to the correct folder depending on whether it passed the checks. Claude then ran the same logic against the remaining invoices in the batch without further guidance.
Both examples share a pattern: a person defines the standard once, either by narrating it or pointing Claude at an existing checklist document, and Claude reapplies that standard to new material going forward.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Is Claude’s record-a-skill feature worth using?
For workflows that involve repetitive judgment calls (checking a list of criteria, sorting files by consistent rules, renaming documents by a fixed pattern) this feature removes a real barrier: you don’t need to write a detailed prompt or design an automation pipeline. You just do the task once.
The tradeoffs are worth knowing before you commit to it:
- File size caps. Files over 10 megabytes aren’t supported yet, which rules out video files and some larger media formats. Document-based and web-based tasks work fine.
- Setup friction. Getting the Chrome extension installed and permissions configured takes a few extra steps the first time, though it’s a one-time cost.
- Sensitivity to noise. Because the recording captures everything on screen, an accidental extra tab or unrelated click can confuse the resulting skill. Keeping the recording focused matters more than it might seem.
- Model independence. The skill itself runs through a browser extension powered by Opus 4.5 regardless of which model you’ve selected in your chat, so switching models won’t change how the automation executes.
For anyone already doing recurring manual checks, be it content QA, invoice sorting, or similar rule-based review work, this is a lower-effort way to hand that off than building a traditional automation.
How does this compare to similar tools?
This isn’t an entirely new category. Browser automation and “show it once” style agents already exist in other tools, including Claude’s own Chrome extension and OpenAI’s Codex. What distinguishes Claude’s Cowork implementation is how little setup it demands: you don’t need to design a flow diagram or chain together steps manually. You demonstrate the task, narrate your reasoning, and Claude infers the rules from the demonstration and any reference documents you show it.
Frequently Asked Questions
Do I need to know how to code to use record a skill?
No. The entire point of the feature is that you demonstrate the task by doing it on screen and explaining your reasoning verbally. Claude converts that demonstration into an executable process without any scripting.
Can I use record a skill in the browser version of Claude?
No. It’s only available in the Claude Cowork desktop app, because the feature requires screen recording, microphone access, and access to local files, none of which the web version supports.
What happens if my files are larger than 10 megabytes?
Files above that size currently aren’t supported, which limits video-based workflows. Document, spreadsheet, and web-based tasks fall well under that threshold in most cases.
Does the model I select in chat affect how the skill runs?
No. Skills that rely on the browser extension execute using Opus 4.5 regardless of which model is active in your chat window at the time.
Can I show Claude an existing checklist instead of listing every rule out loud?
Yes. Pulling up a text document with your criteria and letting Claude read it during the recording is faster than verbally explaining each item, and it produces the same result.