Source: raw/Claude_Unleashes_RECORD_A_SKILL._Here_s_how_to_automate_any_task_with_it..md (walkthrough); raw/x-bookmarks-recent-digest-2026-07-24.md (@claudeai announcement, permalink); raw/Building_Claude_Skills_Just_Got_10x_Easier.md (Ben AI best practices, added 2026-10-28); raw/Anthropic_Just_Solved_Claude_Cowork_s_Biggest_Limitation_Goodbye_MCP.md (Simon Scrapes ManyChat demo, added 2026-10-30) Creator: Eliot Prince URL: https://www.youtube.com/watch?v=p4PYVG6qziM Duration: 15:48 Platform: YouTube Published: 2026-07-22

Record a skill (Anthropic’s label: “teach Claude a skill”) is a Cowork feature launched 2026-07-21. You record your screen while doing a task and talk it through; Claude turns the recording into a reusable skill. Anthropic’s announcement says it lives under Record a skill in the + menu of the Claude desktop app, on Pro, Max and Team plans. Eliot Prince’s walkthrough, published the next day, records a monthly Google Search Console client report and then runs the generated skill end to end through Claude in Chrome and on a schedule. Until this feature, every route to a skill in this wiki started from something written (a prompt, a SKILL.md, a plugin). Here the input is a narrated demonstration.

Dated UI

The walkthrough predates the September 2026 merge of chat and Cowork (Claude Cowork § One Claude). The step “switch the Cowork toggle on” describes the July 2026 two-mode UI. Where the entry point sits after the merge has not been checked.

Key Takeaways

  • What gets captured. The recorder dialog says your screen, clicks, typing and voice are recorded, sent to Claude and turned into a repeatable skill. Eliot’s demo ran about 200 seconds and Claude kept the screenshots and every click.
  • Where it ran (July 2026). Claude desktop app, because the skill may need to drive the computer. Eliot ran it inside his “second brain” Cowork project, though he says any area of Claude works. With the Cowork toggle on, the + button shows Record a skill, then Start recording.
  • Narrate the judgment, not only the clicks. Eliot’s narration covered the clicks (Search Console → property dropdown → Performance → Compare → last 28 days year over year → Export to Excel → save into a client-named folder on the desktop) and also the reasoning: why compare year over year (it removes seasonality), the report shape (five executive points covering the good, the bad and what stayed the same; three actions for the month; a page of commentary), the tone (“keep a positive tone … if there is a problem, talk about the opportunity to fix it”), and the email (short overview, PDF attached, subject “SEO report July”).
  • Claude asks clarifying questions before it builds. In the demo it asked: how should the skill get Search Console data (browser automation, wait for a manual export, or a connector; there was no Search Console connector, so Eliot chose browser automation)? What format for the report (branded PDF)? What should happen with the email (draft only, never send)?
  • Claude picks its own tools. Eliot expected it to drive the browser with Claude in Chrome where no connector exists, and to use an existing connector (his Gmail connector’s draft function) instead of clicking through the web UI.
  • What the generated skill contains. A description, “the rule that matters most” (draft the email, don’t send it), a “what good looks like” section, confirmations it asks for when invoked, a step-by-step workflow that names which MCP connectors and websites to use, and a reference file for branding.
  • You can edit before saving. The skill opens in a side panel. Eliot told Claude in chat to use his existing brand skill instead of generic branding, and the branding reference file was rewritten to say visual identity is not set there and everything comes from the brand skill. Then Save skill, and it appears in the skills area (his was “GSC client SEO report”).
  • Invoke it like any skill: type / and pick it, or ask in plain language (“create me a Google Search Console SEO report for courchevel.vip for July, trailing 28 days”).
  • The run. Claude in Chrome drove the browser hands-free (the screen glows orange while Claude is in control), applied the filters and exported. Output: a branded PDF with call-out blocks and commentary, plus a Gmail draft.
  • Known gap in the demo. The Gmail connector created the draft but could not attach the PDF. Eliot’s proposed fix is to have Claude in Chrome upload it through the Gmail web UI. He did not show this working.
  • Schedule it in plain language. “Turn this into a scheduled task for this website that runs on the 22nd of every month at 10:00 a.m.” Claude created the trigger. The task appears under Scheduled tasks in the left sidebar, and its definition says it runs the recorded skill end to end.

Ben AI’s four practices and an inbox-triage demo (added 2026-10-28)

Creator: Ben AI · URL: https://www.youtube.com/watch?v=zEv0lLYjUHw · Duration: 12:22 · Platform: YouTube · Uploaded: 2026-08-26

Ben used the feature for several weeks with his team before this video (still the pre-merge UI: Cowork tab → + → Record a skill). His argument for it is Polanyi’s paradox (Michael Polanyi: “we know more than we can tell”). Experts can’t fully describe their own process, so a prompted skill misses the edge cases that show up when you actually do the task.

  • 1. Start small. Don’t record a 30-minute process in one go. Record the smallest part that is useful on its own, then add to it or chain more skills.
  • 2. Plan a little. Open every tab and app you’ll use first, so you’re not jumping between unrelated windows. State the goal out loud at the start of the recording (write it down first if needed). Know roughly what you’ll do; rambling and correcting yourself mid-recording is fine.
  • 3. Narrate why, not just what, as if teaching an intern. Ben calls this the most important factor in skill quality.
  • 4. Iterate with the same feature. Run the skill, then record yourself reviewing its run and giving corrections. He says almost every skill needs at least one such pass. The same route works for improving skills built other ways.

His Gmail triage recording, rule by rule. After stating the goal (mark unimportant mail read, flag what needs a reply and draft it in his voice, label, forward), he walked the inbox:

  • Job applicant → draft a thank-you reply saying the application will be reviewed.
  • Paid-collaboration pitches → label “partnerships”.
  • Software notifications and calendar invites → mark read unless something looks urgent (for example an open payment).
  • Inquiries about his businesses → forward to his co-founder.
  • Newsletter replies → read the room: mark read if there’s no question, draft a reply and leave unread if there’s a question or real feedback.
  • Mail from himself (bank statements) and cold sales pitches → mark read; if a cold pitch is a fourth follow-up, draft an “unsubscribe” reply.
  • Subscribed newsletters → label “need to read” and mark read.
  • Finish with a report that puts important emails at the top, schedule daily at 9:00 a.m., run a test now, and never send, only draft.

Claude then asked clarifying questions and chose the Gmail connector over the browser. Ben recommends connecting a connector for the software before recording, since it makes the skill much more efficient than browser use. He says the resulting skill runs daily “and it’s actually working really well.” His wider method for written skills is in Ben AI’s Skill-Engineering Framework.

Simon Scrapes’ ManyChat demo: no API, no MCP (added 2026-10-30)

Source: raw/Anthropic_Just_Solved_Claude_Cowork_s_Biggest_Limitation_Goodbye_MCP.md · Creator: Simon Scrapes · URL: https://www.youtube.com/watch?v=9QSumbWubZ4 · Duration: 12:44 · Platform: YouTube · Uploaded: 2026-07-23

Simon recorded the Instagram DM lead-magnet setup he does by hand for every YouTube upload, in ManyChat, which he says has no API and no MCP server. His take: any app you can click through can now be taught.

  • The recorded task. Find the latest row in his Notion video-transcripts table whose DM-automation column is “not started”; in ManyChat duplicate an existing automation, rename it to the row’s ManyChat keyword (same case), attach the latest Instagram reel, set the trigger keyword, replace the “watch it here” link with the row’s video URL, save, update, go live.
  • Size. About 240 seconds of narrated recording produced 61 captured steps. He thinks the recorder is Mac-only for now (not confirmed).
  • Claude caught his mistakes. Its clarifying questions asked whether to set the Notion row to “live” afterwards (a step he forgot), and why he downloaded the video and thumbnail from Drive without using them (left over from another process, so he told it to skip them).
  • The generated skill. “ManyChat DM automation”, with a description saying when to use it, and steps naming the Notion connector (it found the table URL itself), the fields to read, and the ManyChat clicks.
  • The test. He deleted the automation, changed the keyword to Fable_123, and asked a fresh session to “do this for the latest video” with auto-approve on. It used Claude in Chrome (enable it in Settings first), worked from screenshots (“quite slow” but precise), matched the right reel, set the keyword and link, went live, and marked the Notion row live. He ran it on Fable but says a Sonnet model would do, since it is only following instructions.
  • Next step. Wrap it in a daily scheduled task: if there is a new video, run the skill. He expects heavy token use and accepts it for the time saved.
  • Later view (October). In his October videos he notes the recorded skill came out “surprisingly small” because it captured only the corrections and niche knowledge Claude lacked, which he now sees as what a skill should be (Advice Reset).

When to use it

  • Tasks that are hard to describe in a prompt but easy to perform: log into a portal, click through filters, download, file, draft.
  • Tasks that reach into apps with no connector. Eliot’s second example: logging into seven different client portals to download payment requests or invoices, saving them to Google Drive, then creating invoices and drafting emails.
  • Creator claim: Eliot says client reporting used to cost him “a week every month” when he ran an SEO agency. He suggests recording every task you do in a day and seeing which ones Claude can then run.
  • Prefer a written skill (see the skill-creator coverage in Building Skills Guide) when the task is mostly reasoning or writing with no UI steps. Recording adds value when the clicks are the hard part.

Try It

  1. Pick one boring, repetitive weekly task that involves a browser. Eliot’s advice: “pick one task in your business this week”.
  2. Before recording, write down the decisions you make, not only the clicks: comparison windows, what “good” output looks like, tone, what must never happen automatically.
  3. Record it while narrating both. Answer Claude’s clarifying questions with explicit safety choices: draft only, never send.
  4. Before saving, read the generated skill. Check the “rule that matters most”, the tools it chose, and any reference files. Point it at your existing brand or voice skills instead of letting it invent branding.
  5. Run it once by hand and watch the browser session. Note every step it fails (in the demo, attaching the PDF) and edit the skill.
  6. Only after a clean manual run, ask Claude to turn it into a scheduled task, then confirm the trigger under Scheduled tasks.

Implementation

Tool/Service: Claude desktop app, Cowork; Claude in Chrome for browser steps; any connectors you already have (Gmail in the demo).

Setup: Desktop app → (July 2026 UI) Cowork toggle on → + → Record a skill → Start recording → narrate the task → answer clarifying questions → review/edit the generated skill → Save skill → test → optionally schedule.

Cost: Included on Pro, Max and Team per Anthropic’s announcement. Browser-driven runs use your normal plan usage. No figures were given.

Integration notes:

  • Claude prefers a connector when one exists and falls back to driving the browser. If a SaaS has an API connector, connect it first for a more reliable skill.
  • Recorded skills can reference other skills. The demo’s branding step defers to a separate brand skill (see Brand Print skill).
  • Scheduled runs that drive a local browser likely need the desktop machine on.^[inferred] The Cowork article documents that a cloud task working on local files loses read and write access when the computer is closed. Eliot’s schedule assumes he “turns up at my computer” to review and send.

Open Questions

  • Post-merge entry point. Where does Record a skill live after the September 2026 chat/Cowork merge, now that there is no Cowork toggle?
  • Artifact format. The walkthrough shows a description, a workflow and reference files, which looks like a standard skill folder. Whether it is a portable SKILL.md that Claude Code can use was not shown.
  • Limits. Recording length, whether narration is required, and how sensitive on-screen data (logins, client data) in the recording is stored are not covered by either source.
  • Reliability over time. Browser-driven skills depend on UI layout. How recorded skills behave when Search Console or a portal changes its UI is untested.