The argument in one line.
Claude Cowork's Save Skill feature can watch a single ~140-second screen-recorded demonstration of a recurring task and turn it into a reusable, natural-language-triggered skill, but the AI's first inferred draft still needs a short human refinement pass to add the guardrails a person doing the task by hand would apply automatically.
Read if. Skip if.
- You already use Claude Cowork or Claude Code and have a recurring research or admin task you'd like to turn into a one-line trigger instead of writing the automation by hand.
- You're curious whether Anthropic's new record-a-demo-to-generate-a-skill feature actually works, before investing time learning to write skills manually.
- You want a concrete, working example of an 'AI news to video-idea backlog' pipeline, including the source list and the guardrails that keep it clean.
- You want a from-scratch introduction to Claude Cowork or Claude Code — this assumes you already know the chat/project interface.
- You need English-language instructions — the demo, the generated skill, and the presenter's narration are entirely in Turkish (on-screen text and spoken audio).
The full version, fast.
The creator screen-records himself doing a real recurring task for his channel: browsing AI-news sites (The Decoder, VentureBeat AI, Webtekno) and logging fresh headlines as new rows in a markdown table of video ideas. Claude Cowork's new 'Save Skill' feature watches that single ~140-second demo and drafts a reusable skill on its own — a description plus natural-language trigger phrases in Turkish. The first draft is usable but naive, so he adds three explicit rules: skip headlines already logged, cap each video-idea line to one short sentence, and never pull from paywalled sources. Re-running the refined skill unattended correctly appends three new, non-duplicate rows. The lesson: the feature removes the need to hand-write skill instructions, but it doesn't remove the need for a human to spot and add the guardrails a first-pass AI inference misses.
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
01 · Cold open & preview
Talking-head hook and a quick look at the finished AI-news-to-video-idea tracking sheet before the live demo begins.
02 · Live demo: doing the task by hand
Screen-records himself browsing The Decoder and VentureBeat AI for fresh AI headlines and manually adding rows to the video-idea tracking table — this roughly 140-second clip is what Claude Cowork will learn from.
03 · Claude auto-writes the skill
From the recorded demo alone, Claude Cowork drafts a skill file: a description, natural-language trigger phrases in Turkish, and the logic for scanning sources and appending rows to the video-idea markdown table.
04 · Refining the skill
He adds three guardrails to the auto-generated draft: skip headlines already in the sheet, keep each video-idea line to one short Turkish sentence, and never include paywalled sources.
05 · Testing the refined skill live
Re-runs the skill unattended; it re-scans the same news sites and correctly appends three new, non-duplicate rows to the tracking table.
06 · Verdict & course plug
Wraps with his take on when the feature earns its keep, then points viewers to his Digital Academy Claude Code course and Skool automation community.
Lines worth screenshotting.
- Claude Cowork's Save Skill feature can watch one live screen-recorded demonstration and generalize it into a reusable, triggerable skill — no manual scripting required.
- The generated skill captured semantic trigger phrases in the user's own language ('scan the headlines', 'update the video-idea table'), not just a literal click-by-click replay script.
- A single roughly 140-second demonstration was enough raw material for Claude to draft a full skill description, a condition list, and file-target logic pointing at an existing markdown table.
- The first auto-generated skill draft needed manual refinement: the creator added three guardrails — skip duplicate headlines, cap the video-idea text to one sentence, and never include paywalled sources.
- Once refined, the skill correctly scanned three real news sites unattended and appended exactly three new, non-duplicate rows to the tracking table on a live re-run.
- The demo-to-skill loop turns a recurring content-research chore into a one-line command: do it once by hand, let Claude infer the workflow, patch the gaps, then just ask for it by name from then on.
- The creator's own saved-chat history (visible in Claude Cowork's home screen) shows he already automates other recurring business tasks — lead triage, engagement boosting, invoice chasing — so the Skill feature slots into an existing automation habit rather than a one-off experiment.
- Guardrails obvious to a human performing a task by hand (don't re-log a headline you already have) are exactly the kind of detail an AI's first inferred version of that task tends to miss.
One demo, then a skill — but the guardrails still need you
Anthropic's Save Skill feature turns a single screen-recorded demonstration into a reusable, natural-language-triggered workflow, but the AI's first inferred version of that workflow still needs a short human pass to add the guardrails that keep it from repeating mistakes.
- The training material Claude needs isn't written instructions — it's a single recorded pass of you actually doing the task, browser and all.
- He kept the demo short (about 140 seconds) and used real destinations (The Decoder, VentureBeat AI) rather than a staged toy example.
- From one demo, Claude drafted a full skill description plus a list of natural-language trigger phrases in the user's own language, not just a literal click-replay script.
- The auto-generated skill file targeted a specific existing artifact (a markdown table already in use) rather than inventing new output structure.
- The first draft was usable but incomplete: it took a manual pass to add the three rules that actually prevent bad output — no duplicates, no paywalled sources, and a length cap on the generated text.
- Guardrails that are obvious to a human doing the task by hand (don't re-add a headline you already logged) are exactly what an AI's first inferred version of a skill tends to miss.
- Re-running the refined skill unattended produced exactly three new rows with no duplicates — the guardrails from the refinement step held up on a real run, not just on paper.
- Confirming a skill actually works means re-running it against live, changing data (fresh headlines that day), not just checking that the generated instructions read correctly.
Terms worth knowing.
- Claude Cowork
- Anthropic's collaborative AI workspace product, distinct from the plain Claude Code CLI, with chat, project folders, and — as shown in this video — the ability to record a demonstrated task and turn it into a reusable skill.
- Skill (Claude Cowork)
- A saved, reusable set of instructions and trigger phrases that Claude follows automatically when a matching request comes in. It can be written by hand or, as demonstrated here, generated from a recorded demonstration of the task being performed once.
- The Decoder
- An English-language AI and tech news outlet the creator monitors as one of his sources for video-idea material.
- Skool
- A community and course-hosting platform. The creator runs his paid automation community there, referenced at the end of the video as a call to action.
Things they pointed at.
Lines you could clip.
“The Decoder, VentureBeat AI, Webtekno vb.'i tarayıp güncel yapay zeka haber başlıklarını video-fikri.md dosyasındaki tabloya yeni satır olarak ekler.”
“Sadece sayfada zaten olmayan başlıkları ekle. Video Fikri alanını kısa tut, Türkçe tek cümle. Ücretli/paywall arkasındaki içerikleri asla ekleme.”
Word for word.
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
The bait, then the rug-pull.
The video opens with a promise pulled straight from its title: writing Claude skills has 'changed forever.' The proof is a live screen recording — a real, recurring content-research task performed once by hand, then handed to Claude Cowork's new Save Skill feature to see whether the tool that watched him work could write the automation itself.
Named ideas worth stealing.
Demo → Skill → Refine → Reuse loop
- Record yourself doing the task once
- Claude drafts a skill from the recording
- Add the explicit guardrails the draft missed
- Trigger the finished skill by a natural-language phrase from then on
The four-step loop Claude Cowork's Save Skill feature runs on, demonstrated live in this video.
How they asked for the click.
“Points viewers to the Digital Academy Claude Code course and the skool.com/otomasyon community”
Soft close after the demo — screen-shares the course/classroom tiles rather than a hard verbal pitch; the sparse audio transcript couldn't confirm his exact spoken words here.
Visual structure at a glance.
