10:57Brock Mesarich | AI for Non Techies · Review
Claude Cowork Comes to Mobile and Web
A hands-on first look at Claude Cowork for mobile and web, and the one-way sync limitation Anthropic didn't mention.
July 9thThe argument in one line.
Claude Cowork's Record a Skill feature turns a narrated screen recording into a working skill file, but it only earns its place for tasks in apps that have no native connector — everything else is still better served by a direct integration.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Cowork or a similar AI assistant and want to automate a recurring browser task that has no official connector or MCP integration.
- You manage a YouTube channel, community, or content pipeline and manually triage comments or similar recurring admin work.
- You're deciding whether to invest time recording a skill versus building a Zapier/MCP connector for a workflow.
SKIP IF…
- Every tool in your workflow already has a native connector — this feature adds nothing over connecting directly.
- You're not using Claude Cowork specifically; this is a Cowork-only feature, not general Claude or the API.
The full version, fast.
Anthropic added a Record a Skill feature to Claude Cowork: you screen-record yourself performing a task while narrating it, and Claude turns the recording into a written skill file it can replay later. The creator tests it on a real task — pulling YouTube comments into a tracking spreadsheet and deciding which ones justify a video idea — and Claude correctly reverse-engineers the steps, including an unstated rule (never touch the comments) it inferred from watching. His verdict: this is most valuable for apps with no existing connector, where browser-use alone tends to fail; if a connector or Zapier MCP integration already exists for the app, use that instead since it's more reliable than replaying a recorded walkthrough.
Free for members
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.
00:00 – 00:47
01 · Cold open + promise
Creator explains that Anthropic added a new way to build Claude skills: record your screen, talk through what you're doing, and Claude turns it into an automated skill. Says his verdict on whether it's a game-changer will surprise viewers.
00:47 – 01:51
02 · Opening Claude Cowork's Record a Skill
Shows that this feature lives inside Claude Cowork's desktop app, accessible via the plus button once Cowork is toggled on. Clicking 'Record a skill' shows a warning not to record passwords, secrets, or private conversations while recording.
01:51 – 02:58
03 · Recording the demo task
Creator starts the recording and narrates the task live: pulling YouTube comments (only long-form videos, not Shorts) into a 'YouTube Comment Tracker' spreadsheet, and judging whether each comment is worth turning into a video idea.
02:58 – 03:53
04 · Walking through YouTube Studio and the sheet
He filters Community comments to unresponded/long-form, copies example comments, pastes them into the tracker with text wrapping, and demonstrates picking Yes/No on whether a comment merits a video idea.
03:53 – 04:13
05 · Ending the recording
Clicks Done; Claude begins processing the roughly 73 steps and about 11 minutes of recorded demonstration plus the full narration transcript.
04:13 – 05:05
06 · Claude's generated skill summary
After about two minutes, Claude returns a plain-language breakdown of the skill: grabs comments from YouTube Studio with the right filters, logs them into column A of the tracker sheet avoiding duplicates, and makes the Yes/No call based on the examples shown — plus a rule it inferred on its own never to touch (reply/like/delete) the original comments.
05:05 – 06:15
07 · Reviewing the SKILL.md file
Creator opens the actual generated skill file, showing it names the tools used (Claude in Chrome MCP), documents step-by-step navigation, describes the tracker's column layout, and defines what qualifies a comment as a video idea plus edge cases. Saves the skill for reuse.
06:15 – 06:40
08 · When this feature actually matters
Creator's core verdict: record-a-skill mostly pays off for tasks with no existing connector. If an app is already connected (Slack, Gmail, Google Drive, etc.), that direct connector is more reliable than recreating a walkthrough.
06:40 – 08:00
09 · Browser use vs. recording a skill
Explains what 'browser use' is — Claude driving a real browser click-by-click — and why relying on a cold prompt for a multi-step, app-specific task is less reliable than talking Claude through it live, the way you'd train a new employee.
08:00 – 08:48
10 · The Zapier MCP alternative
For apps with no Claude connector at all, the creator recommends Zapier MCP, which connects Claude to 9,000+ apps. He shows his own Zapier MCP server, highlighting Skool and Beehiiv — tools he uses daily that have no native Claude connector — and says he'd default to that over recording a skill.
08:48 – 09:35
11 · Privacy note + closing take
Flags the privacy tradeoff of letting Claude watch your screen (avoid recording banking info or anything confidential), reiterates this is a step toward AI that learns tasks by observation rather than instruction, and closes with a subscribe CTA.
Lines worth screenshotting.
- Claude Cowork can now watch a narrated screen recording of a task and turn it into a written, reusable skill file — no manual skill-authoring required.
- The recorded skill correctly inferred an unstated rule (never reply to, like, or delete the source comments) purely from watching the demonstration, without being told.
- The creator's rule of thumb: record a skill only for apps with no existing connector — if a direct connector or MCP integration exists, use that instead.
- Browser-use and computer-use features tend to fail on multi-step, app-specific navigation because a single text prompt can't specify every click and screen location.
- Narrating a task while recording is closer to training a new employee than prompting an AI, and that difference is what makes the recorded skill more reliable than a cold browser-use prompt.
- Zapier MCP connects Claude to 9,000+ apps including ones with no official Claude connector, such as Skool and Beehiiv, which the creator says he'd use over recording a skill when available.
- The full skill-recording and processing cycle for an 11-minute demonstration ran through roughly 73 recorded steps and took about two minutes for Claude to convert into a finished skill file.
- The feature is currently exclusive to Claude Cowork's desktop app, not the standard Claude web or mobile app.
Match the automation method to whether a connector exists
WHAT TO LEARN
A new AI feature that learns tasks by watching you is genuinely useful, but only for the specific gap it fills — everything else still has a better tool for the job.
- Before automating any recurring task, check for a native connector first — a direct integration will always be more reliable than a workaround, however clever.
- When no connector exists but the task lives entirely in a browser, narrating a live demonstration gives an AI far more usable detail than a single written prompt ever could.
- An AI system watching a full demonstration can infer unstated rules and constraints (like 'never touch the source data') that you'd otherwise have to spell out manually.
- Third-party integration layers like Zapier MCP can cover the remaining gap for apps that have neither a native connector nor a practical browser-use path.
- Any tool that records your screen and voice carries a real privacy tradeoff — treat what you say and show while recording the same way you'd treat a live screen share with a stranger.
Terms worth knowing.
- Claude Cowork
- Anthropic's desktop app mode for Claude that adds project/task management and screen-recording features on top of the standard chat interface.
- Record a Skill
- A Claude Cowork feature that captures your screen, clicks, typing, and voice while you perform a task, then converts that recording into a written, replayable skill file.
- Skill file (SKILL.md)
- A markdown document Claude writes describing the tools, steps, and rules needed to repeat a task, created either manually or from a recorded demonstration.
- Connector
- A built-in integration inside Claude that lets it directly read and act inside another app (e.g. Gmail, Slack) without needing browser automation.
- Browser use / computer use
- Claude capability that lets it operate a real browser or computer by clicking, typing, and reading the screen like a person, used when no direct connector exists.
- Zapier MCP
- A hosted MCP server from Zapier that connects Claude (and other AI tools) to 9,000+ third-party apps, letting Claude take actions in apps that have no native Anthropic connector.
Things they pointed at.
Lines you could clip.
06:15
“If there's a task you wanna do that has a connector, I would consider using that first. If it doesn't have a connector, you can then record and walk through a skill using the browser use capabilities.”
07:56
“I think it's much more valuable in order to screen record what it is you're doing and literally talk to Claude in real time about how to perform this task rather than just sending off a prompt to Claude using browser use and expecting it to know exactly where to click on things on your computer.”
Word for word.
Read-along
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.
analogy
00:00Anthropic just dropped one of the most interesting features that I've ever seen for Claude, and this might completely change the way that we use it. This feature is a new way to build out Claude's skills. So instead of needing to write skills, we can now have Claude record our screen, watch what we're doing, and we could physically talk through the process of what we're wanting Claude to automate, and then it's gonna analyze that and turn that into a skill.
00:22So in this video, I'm gonna break down exactly how it works. I'm gonna test it out so you could see it in action, and then we're gonna break down whether or not this is a game changing feature like most people are gonna be talking about. And I will say my answer is probably gonna surprise you.
00:36So without further ado, let's dive right into this video. Alright. So without wasting any time, let's dive in and begin testing this out.
00:42So first of all, this is only available inside of Claude Cowork. And if you wanna use Claude Cowork, make sure to download the desktop app. There will be a link in the description to do so.
00:52And now where we're gonna see this is if we come over to this plus button here, once we toggle on co work, then we will now see this little button here that says record a skill. If I click on this, what it's gonna do now is it's going to give me this prompt here. Right here, it says record a skill.
01:09Your screen clicks typing, and voice are recorded, then sent to Claude and turned into a repeatable skill. Don't type passwords or secrets or display sensitive information or private conversations while recording. And this brings up a really good point, and that is security, which is something I'm gonna talk about later on in this video and give you my honest thoughts on.
01:25Next up, we have two other buttons. We have a start recording button, and then we have our microphone because we're able to literally talk through while we are doing something on our screen as if we're teaching an employee how to do a specific task or process. Let's test this out now.
01:39So I'm just gonna click on record. Next step, you could see it pulled up this little icon here, and now it's actually recording my physical screen. So let's go and explain a task.
01:48Alright. So what I am trying to do is I want to automate, basically, copy and pasting different YouTube comments and adding it to this YouTube comment tracker spreadsheet.
01:58From here, I basically wanted to decide whether or not it's worthy of creating a video in order to give me YouTube video ideas based on the comments that I'm getting from my subscribers. So let's come here, and I'm just gonna physically show on my screen how to do this task. Because before, if we were to create this skill, then I would have to, like, explain to a tee which buttons to click on, where on the screen to click, etcetera, for it to understand how to navigate through YouTube Studio, which is what I'm having it go through.
02:26And that could be a bit of a headache, and that's kind of where the browser use and computer use capabilities of things like Claude and Claude Cowork kind of fall short. So, this fixes it. Anyways, I'm just gonna come here.
02:37So, Claude, what I want you to do is I want you to take all these comments that I get for only my long form YouTube videos, and I want you to paste in the comments. From here, I want you to, for all of these, make sure that you are wrapping the text so that way, you know, it shows the full comment there. And then for each of these, I want you to analyze based on the video.
02:59You know, is this worthy of a video idea? And for this one, for example, what's the cost of the entire project that's probably not worthy? So I want you to go through in reason based on what you know about my YouTube videos and the ones that perform well in order to decide whether it's worthy of making a video about.
03:14And, again, I want you to only go through my, like, long form YouTube videos. Don't go through shorts. What I would do is I'd come here and select just my videos for content type and then copy all these.
03:24So for example, would be good to get a side by side co work versus, um, Chatcha b t work competition. Just open up the spreadsheet, paste that in. Again, you wanna make sure that it's always wrapping the text.
03:35And then this one, in my opinion, is actually worthy of a video idea. So this is the process. I just want you to go through, copy all of these comments, and then basically decide whether or not this is worthy of a video idea.
03:45This one, I would say, isn't. Alright. So now I'm just gonna click on done, and you could see it is actually processing the 73 steps that it saw through the video.
03:55And this is my first time actually testing this out, so I'm really interested to see how capable and how spot on this actually is. And I'm interested to see the logic behind how it's able to break all this down.
04:06So looks like it took screenshots throughout this entire workflow, and I believe it took about 76 screenshots or something like that.
04:14And then it's also transcribing the entire recording of the video right here. This is the full transcription.
04:19Okay. So now it took maybe, like, two minutes here, and then it came up with this. So it went through, watched the recording, and turned it into a skill.
04:26Here's what it does in plain terms. Then I'm gonna read through very briefly this skill and the file that it created here that I could then save as a skill. So first of all, it says it grabs the comments, opens YouTube Studio, sets the filters like I did because I explained that in my voice over.
04:40It then logs them into the spreadsheet, pastes them into column a. You can see it's pretty descriptive of exactly, like, where exactly we're pasting this information, what buttons to click on, etcetera. Then it makes a call, picks whether or not it is good for a YouTube video.
04:54And then it also noticed, and I never mentioned this, that it never touches the comments, So it doesn't reply, like, or delete these comments, which is cool that I was able to figure that out itself, which I guess is kinda self explanatory. And then now we could actually save this skill here. But, actually, let me go ahead and save that.
05:09Now here we go. Let's quickly just go through the actual scale MD file that it wrote so we could see basically how detailed it is, which is obviously gonna be really important to recreate that workflow. So it breaks down what exactly this is, turns viewer comments into a video idea pipeline, and it breaks down the specific tools that it's gonna use, which it's using the Claude in Chrome MCP tool.
05:29It breaks down step by step how to do this, breaks down how to log them into the tracker, and even shows a visual of what the columns look like. And then here is a pretty detailed breakdown on what justifies a video idea and when it should select either yes or no, and then breaks down the report back here as well as some specific edge cases.
05:47So this is a pretty detailed skill and d file, and, obviously, we could update this over time. Alright. So now that we saw it in action, let's break down my actual thoughts on this, if I think it's capable, when I think we should use this, if we should even be using it, and other alternatives that might be better than using the record skill feature that they just released.
06:04First of all, what I think after testing this out is I think that the majority of the time that we'll be using this skill record feature is for different things that don't have specific connectors inside of Clot. If you don't know what a connector is, this is how it's able to work across our different applications like Slack and Gmail, Google Drive, etcetera.
06:22It's an MCP server that it's using in order to navigate throughout those applications. I do think that that's gonna be more reliable than having to walk through and try to basically explain how to do different tasks throughout these applications while screen recording. But where I do find this very valuable is when we need to do things that require browser use capabilities.
06:40If you don't know what browser use, this is basically the function that allows Claude to go and actually use our browser or our computer, do things like fill out forms, click on buttons, navigate different websites, etcetera. So more of the manual work that you actually do on your computer aside from using a connector.
06:56And a perfect example of that is what I showed you inside of the YouTube app. There is no way to connect Claude to YouTube, so I have to instruct it manually in order to go and copy those comments and paste them into Google Sheets for me. And I think it's much more valuable in order to screen record what it is you're doing and literally talk to Claude in real time about how to perform this task rather than just sending off a prompt to Claude using browser use and expecting it to know exactly where to click on things on your computer.
07:23So to summarize that, if there's a task you wanna do that has a connector, I would consider using that first. If it doesn't have a connector, you can then record and walk through a skill using the browser use capabilities. Now if there is an app that you can't connect to inside of Cloud, I highly recommend to use something called the Zapier MCP.
07:40If you don't know what this is, Zapier is a platform that connects to 9,000 plus different applications, and it's really easy to plug Cloud directly into all those apps. It includes the majority of the apps that you use on a day to day basis. Now in order to give you an example, this is an MCP server inside of Zapier that I created, and there's a bunch of different apps that I could use that I can't connect to inside of Claude.
08:00So for example, school and beehive are different tools that I use every single day to help grow my business, but there's no way to connect those inside the Cloud app. Well, with the Zapier MCP, I'm able to perform actions across both of those different applications, and I personally would prefer to use this as opposed to using this new record skill feature from Anthropic and hoping that the browser use is able to figure out how to do those tasks.
08:22If you guys do wanna try out Zapier MCP, there is a link in the description for you to test it out. Now I do wanna say I really do think this is a step in the right direction. I personally know friends of mine in the AI space that have meta glasses that are literally recording what they're doing on their computer all day long in order to help them optimize their workflows and come up with ideas for what Claude can help them automate.
08:43And I think this is a step in the direction for us to be able to, like, record our screen, walk through as if we're talking to an employee on how to do a specific task, especially ones that are a little bit more intricate and require a little bit more attention to detail instead of us needing to go and just, like, type in a prompt and hoping Claude can understand what we're trying to accomplish.
09:01Now the question you probably have is, is this safe based on privacy? And if you are worried about privacy, then maybe don't use this. You might not want Claude to see what it is you're doing on your computer and have access to all this information.
09:11If you do wanna test it out, just make sure not to show, like, any banking information or anything that's confidential you don't want Claude to have access to. Other than that, I personally have no problem testing this out for myself if I'm not showing any confidential information. Anyways, guys, I hope you got some value.
09:25If you guys want more videos like this, subscribe to the channel. I cover all things AI for nontechies and have countless videos on how to use Claude Cowork. With that being said, thanks for staying the end, and I'll see you in the next video.
The Hook
The bait, then the rug-pull.
A new record-your-screen feature promises to turn any narrated demonstration into a reusable Claude skill — no prompt engineering required. The creator puts it to a real test before deciding whether it lives up to the hype.
Named ideas worth stealing.
06:15concept
Connector-first decision rule
- Has a connector? Use it directly.
- No connector, but it's browser/manual work? Record a skill.
- No connector at all, and Zapier supports it? Use Zapier MCP instead.
The creator's three-tier rule for choosing how to automate a task in Claude: prefer a native connector, fall back to recording a skill for pure browser-use tasks, and reach for Zapier MCP when an app has no Claude connector but is supported by Zapier.
Steal forDeciding how to automate any recurring manual task inside an AI assistant without over-investing in the wrong integration method
How they asked for the click.
VERBAL ASK
09:15subscribe
“If you guys want more videos like this, subscribe to the channel.”
Simple end-card subscribe ask paired with a closing diagram summarizing the feature; sponsor plugs (Zapier MCP, Claude Cowork download link) live in the description rather than a hard mid-roll pitch.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
