Most AI automation still demands too much setup: hand-crafted prompts, Zapier chains, or actual code. Claude’s new skill recording feature sidesteps all of that. You demonstrate a task on screen, narrate what you’re doing and why, and Claude packages it into a repeatable skill you can trigger any time.
It’s the difference between writing a recipe from scratch and just cooking while someone watches over your shoulder.
What Skill Recording Actually Is
Inside the Claude desktop app, under the + menu, there’s a “Record a Skill” option. Hit it, and Claude starts watching your screen and listening through your microphone. Walk through your task once—click around, explain your logic out loud—then stop the recording. Claude analyzes what it saw, codifies the steps, and saves the result as a named skill tied to your account.
From that point on, you invoke it with a slash command (like /invoice-check or /content-qa) and Claude executes the full workflow on its own.
A few things to know before you start:
- Desktop app only. This isn’t available in the browser version of Claude.
- Enable code execution and file creation under Settings → Capabilities, or the skill won’t be able to act independently.
- File size cap of 10 MB. Document and spreadsheet workflows work great; anything involving large media files won’t.
- Keep your screen clean. Claude ingests everything it sees. Wandering into an unrelated tab mid-recording adds noise and can muddle the resulting skill.
Two Workflows Worth Building Right Now
Quality Assurance Checklists
If you regularly review work against a fixed list of criteria—published blog posts, client deliverables, product listings, anything—this is the fastest win.
Say you manage a small e-commerce store and every new product page needs to have: a title under 70 characters, at least three product images, a filled-in meta description, a price, and a shipping weight. Right now you probably check that manually or paste a checklist into a chat each time.
Instead: pull up your checklist document, start a recording, explain what each criterion means and why it matters, then walk through checking one product page yourself. Stop the recording. Claude builds the skill.
Next time a batch of new listings goes live, you run the skill and get back a structured report: which pages passed, which failed, and exactly what’s missing. Schedule it weekly and you’ve got an automated QA pass running without you.
One practical hack: if your criteria are already written down somewhere—a Google Doc, a Notion page, a plain text file—just have it open during the recording. Claude can read it directly. You don’t need to recite every bullet point out loud.
Document Processing and Filing
Repetitive file-handling tasks are another natural fit. Think about any process where you open a document, check it for completeness, rename it to a standard format, and move it to the right folder. That’s exactly the kind of thing that eats 20 minutes a day in aggregate but feels too small to formally automate.
A concrete example: you receive contractor invoices in a shared folder. Each one needs to have a vendor name, invoice number, date, line-item totals, and a tax amount. Valid ones get renamed to a standard convention (YYYY-MM-DD_VendorName_INV####) and moved to a processed folder. Anything incomplete or suspicious—mismatched totals, missing fields, a duplicate invoice number—gets routed to a needs-review folder for you to look at later. Then the data from valid invoices gets logged into a tracking spreadsheet.
Record yourself doing that once for a single invoice. Claude handles the rest of the batch.
How to Record a Skill That Actually Works
The recording quality determines the skill quality. A few habits that make a real difference:
Narrate your decisions, not just your actions. Don’t just click—say why. “I’m moving this to needs-review because the subtotal and total don’t match” gives Claude the judgment logic it needs, not just the mechanical steps.
Use a real example, not a toy one. Record against an actual document or page, not a placeholder. Edge cases show up in real work, and Claude learns from seeing them.
Keep the demo tight. A 60-to-90-second recording of a focused task outperforms a five-minute ramble. If a task has multiple distinct phases, consider breaking it into separate skills.
Prepare any reference material in advance. If you’re working from a rubric, checklist, or naming template, have it visible on screen before you hit record.
The Broader Shift This Represents
The hard part of AI automation has always been translation: taking what’s in your head and turning it into something a machine can execute. Prompts help, but they require you to think like a programmer. Visual workflow builders are powerful but have steep learning curves.
Showing is different. Most people can demonstrate a task in under two minutes, especially if they already do it regularly. That demonstration becomes the spec.
The real unlock here isn’t any single workflow you build—it’s developing the habit of asking, whenever you do something repetitive: could I just record this? Once that reflex kicks in, you’ll find automatable tasks everywhere.
Start with the thing you do most often that you also trust yourself to define clearly. Record it once. See how close Claude gets. Iterate from there.
