Stop Prompting Claude. Press Record Instead.

Anthropic just released a new feature that records your work and automates exactly what you're doing. So, in this video, I'm going to break down exactly how it works, the non-obvious ways to use it, and then the ways to not use it. And at the end, I'll dive into something that 99% of people miss, but is the real reason this tool will transform your AI productivity. By the way, I'm Austin. I'm not just a content creator. I ran an engineering team at JP Morgan. I was a COO of a tech startup with over 25 million. And now I work directly with top AI companies like Anthropic. So, part one, what it is and how to use it. So, this new feature lets you record your computer's screen as well as its audio. And it'll take that information, analyze it, and create a skill, which you can then use to automate the process. Now, before we get into how to use it, let's see it in action because that'll just help put it all together. So, I'm going to automate my email archiving process. So, I go to Claude Co-work and then I hit the plus and select record a skill. I go to my email and then rightclick, hit archive, and then I say I'm archiving things I don't need to read. If it looks like it was sent from a robot or cold outreach, archive it. I hit stop and it starts analyzing my session and then from there it'll create a skill that will complete the task for me. Now throughout the video if I talk about recording a skill, this is the process that I just went through. When I say automating something, I'm referring to the process of taking that skill and creating a schedule so that it'll run without you. And this is really as simple as just saying run this skill every day at 8 a.m. and it will create a scheduled [music] task. Now, this is a pretty straightforward tool, but before I get into how to use it and when not to use it, there are two highlevel observations that you need to understand. The first is that Claude is trained to understand the intention of the user, not just what the user did. And by taking that intention, it can reverse engineer an optimal solution for a computer to complete the task, not necessarily a human to complete the task. So in the email example, I click buttons on the Gmail website, but Claude has access to the Gmail API, which allows it to programmatically do this exact same task, but instead of going through the website, it uses the API. The beauty here is that I never said anything about using an API during the recording. It just knew to suggest this based on what I was doing. And as somebody who used to build businesses entirely around automating clicking buttons on the internet, I know firsthand how fragile this can actually be. So for it to be smart enough to suggest a more stable, streamlined, and programmatic approach. This is extremely impressive and helpful. It's not just doing what you do, it's doing level two thinking to come away with a better solution. The second is that you're able to piece together videos as you build out automations. So, let's say you wanted to add some of that workflow that I just made, for example, send a Slack message after archiving the emails. I can just record my screen again and it will enhance the skill it just created and then add an additional step. Okay, so that's how to use it. But this is not as simple as just recording everything you do and then just telling Claude to go because the reality is that won't work and will cost you a lot of time and money. So, there are four things that you have to do. The first is that you have to narrate why you click things, not just what you clicked. This may feel clunky the first time you do it, but this is an absolute requirement. As you record the task, you have to add context as to why you're doing exactly what you're doing, not just about what you're doing. The best way to imagine this is imagine you're explaining how this process works to a curious 12-year-old. They would ask you, "Wait, why did you do that and why didn't you do that? Can I click that button? Can I do this? Why should I do that?" Imagine you're getting all of those questions while you're recording, and you have to answer them on the fly. So, you wouldn't just say, "Go to Gmail, right click, and then hit archive." That's actually pretty useless because it sees that in the recording. Instead, what you should say is, "This first email I didn't archive because it is clearly an email from a client and I have yet to respond. The second email I did archive because it's a newsletter I already open. And the third you can archive as it looks like no action is needed." This is the level of context that helps create effective automations that work in the real world, not just in sexy demo videos. Be clear with why you did and didn't do things, filling in gaps that the model can use to create the skill itself. Okay, so that's the first of four reasons. The second is something that almost nobody is doing. Record a rubric, not just a task. Most people record themselves completing work, which is powerful, but the unlock is recording yourself reviewing final outputs. Highlight what's good, highlight what's wrong, and why. This lets Claude capture your judgment and your taste, not only your process. And I'm going to dive deep on this later in this video because this is what I see is the most powerful component of this entire tool. Now, before we get to the next two ways on how to properly use this feature, we're covering how audio plus video makes this tool an absolute gamecher. And another place audio can be a gamecher is when you capture the recordings from your meetings, which brings us to today's video sponsor, Granola, which I am absolutely hyped about because this tool has streamlined my entire sales process directly in Claude Code. So, how it works is that Granola transcribes straight off my computer's audio. So no bot has to join the call. And its MCP connector pipes everything automatically into Claude. And this means I only reference these call transcripts when I have to. And one of my favorite examples is working on identifying gaps in the services I provide clients. I can go directly in Claude code and ask what objections came up most this month during sales calls. And from there, it will analyze all of my sales calls and create a short list of these objections. Once I get that, I'll often say something like when and who had these objections and it looks at the history which is stored in granola and provides me with that contextual information. But that MCP pipeline is why it's now part of how I actually run my business. By pairing these two features, it makes my whole workflow seamless. I first came across Granola because my business partner was using it and I was like, "Yo, that's fire." And you guys know that I love this tool. So, I'm excited to partner with them and offer my community 100% off their first month. Click the first link in the description to try it out and you're going to be happy you did. The third step is you need to know why you're automating it and how you do it. So, a couple things here. You need to know why you're automating something. Don't just be an automation junkie. The reality is the fastest way to run out of tokens is by having a 100 scheduled tasks that are just entirely useless. So, know what you're automating, why you're doing it, and when it comes time to actually do the task and record it, be precise with what you're doing. Don't just fumble around on the recording trying to figure out the task. By being precise, this makes AI's job easier and more token efficient when analyzing how to automate the task itself. Also, an unintended side effect of this is it makes you really think about the task and the process at hand, which flexes your critical thinking muscle. And the fourth must do is build small pieces and then an orchestrator. Don't record one 10-minute video doing a million things. Think in compartmentalized tasks that can be atomic skills. Then you can take those atomic skills and put them together to complete the task end to end. When I explain this to people, they'll usually get stuck on where does one task or skill end and when does the next start. And this is definitely an art, not a science. But I like to think of every automation as a series of inputs and outputs. And I try and break it up at each unique step in the process. So in the email example, task one was archiving the emails where the input was nothing. This was the start of the task. And then the output is the total amount of emails archived. So that's the input and output. Then the next step is sending the Slack message. The input is the amount of emails archived from the previous step and the output is the Slack message. So to make this full automation, I would break it up into two atomic skills. The email message and then the Slack message. Now this automation is relatively simple. So you can just do it in one recording and say create two unique skills for each task and then a single orchestration skill that runs it entirely. So this way you just do one recording and then you use Claw to break it up into multiple skills. But for more lengthy and complex ones, I would try and break up each specific step in the process as its own unique recording. And an added benefit of this is when you do break up things into atomic skills, they can be leveraged in different workflows. So for example, if I wanted to send a Slack message in a different automation, I can now reuse that atomic skill I just created. Skills are the building blocks of the broader automation. So these are all best practices. And if you're thinking about tasks to actually get started on screen, you'll see a complete list of good candidates. Now, before you go and start recording everything, unfortunately, there are a ton of ways that you shouldn't use this tool. And there is a cautionary tale from GM that wasted $45 billion automating the wrong things. Part three, how not to use it. So, there's a famous story about how GM made a $45 billion mistake, and you're likely making the same mistake with AI. In the 1980s, GM tried to automate their factories with robots. They spent $45 billion, and it failed. Problem was that they were automating broken processes. They just made bad work happen faster. And today, the same thing is happening with AI. MIT put out a report last year that found 95% of enterprise AI pilots produced no measurable return. Not because the models are bad, but because companies and people are pointing them at workflows that are messy, undocumented, and inconsistent. Much like exactly what happened to GM in the 1980s, it's the same mistake 40 years later. Toyota was doing the opposite in the 1980s and was able to successfully automate processes using what's called the Shingo model. Not to go too deep on the Shingo model, but to summarize, automation only works when the process underneath it is stable, transparent, and manually perfected first. Each of these three words is critical for knowing when not to use this feature to create skills that automate tasks. So, the first word, stable. Only record a process when you know that process isn't constantly changing. And change can come from two different angles, external and internal. Externally, if the system you're using changes all the time, it will be hard to automate anything inside it. For example, imagine a team is actively rebuilding an internal analytics dashboard. That tool is an unstable external system. So building a repeatable solution on top of it just isn't worth exploring. So that's externally. Now internally, the tools you're using may stay the same, but the rules behind the process may keep changing. So maybe you don't know which metrics matter, who approves the report, what the final output needs to include. So if you start creating automations that aren't perfectly fleshed out, you're going to forget about them and it's not going to produce what you actually need. And this disconnect from the automation output and what's actually needed creates operational debt that stacks up fast if you're not careful. The second word is transparent. Only record a process when you can explain exactly how decisions are being made. Transparency can break in two places, steps and judgment. Steps are what needs to happen and in what order. If someone else couldn't follow your instructions manually, AI probably won't follow them consistently either. Judgment is all those invisible rules living inside your head. My email skill works because I can articulate an exact rubric. But most decisions we make all day aren't that clear. We know the right answer when we see it, but we couldn't explain it when someone asks us. And transparency doesn't necessarily mean simple. It means explainable. If you can't say the decision out loud, don't press record. The third is manually perfected. If you've watched my videos, you know a concept I talk about all the time is you have to earn the right to automate. Before AI can repeat your workflow, you need to manually test it, work through edge cases, and then prove it consistently produces the correct result. Now, creating the output is only half the process. Depending on what you're working on, you may need a reliable way to verify it. So, I was doing some research and I somehow found a cautionary tale from a Japanese journal that I thought was perfect. So, shout out my AI agents that help with the research. So, a Japanese consultant recorded himself building an expense summary. And in the sheet itself, there were live formulas like sum if and count if. Shout out all the Excel junkies out there. And when he ran the skill on fresh data, it didn't use the formulas. It calculated the numbers and typed them in as static text. So, I created this Excel sheet and the file looked right, but functionality it was wrong because the formula that allowed the file to get updated periodically was completely gone. So, visually it looked right, but functionally it wasn't. So for a record feature, yes, you can use it to create a skill before you automate it, i.e. create a schedule that runs it for you that I covered earlier in the video. You have to test this skill manually before you do any sort of scheduling or automation so you know for certain that it works. Now, I've gone deep on my channel on how to optimize these skills that you create, which I'll link at the end of this video. Now, before we get to what 99% of people miss, we need to go through our anti-slap agreement. visuals, the testing, the production that went into this video is entirely human-made. These aren't AI generated assets. It's made by humans for humans. So, all that I ask part of this agreement is that you hit subscribe so that this content can reach more people. Also, every video I give away a CloudMax subscription. So, this video's winner is Mloud03, who is a Gen Xer who is using AI to help market a book he wrote. You are an absolute legend. Shout out all the Gen Xers watching. Now, to enter this video's giveaway, comment below with a topic you want me to cover, and I'll also try and respond to every single one of you. Part four, what 99% of people miss. When people see this new feature, their brain immediately goes to completing tasks. But completing a task and creating an output is only half of the process. And honestly, it is the easier part. What excites me is using it to evaluate the output. And if you can successfully evaluate Claude's output, you can ensure the work is high quality before you ever look at it. So, I like to think about training AI like you're training human teammates. Would you onboard somebody and explain everything through a giant written document? Or would you hop on a call, share your screen, and walk through the work? Every time I try and train someone over messages, I quickly say, "It this just isn't as effective if we spent 15 minutes on a call." And it's the same thing with AI. That's why the pairing of the visual with the audio is so exciting. You can really explain why something is good and why it isn't. So, here is the three-step process I use when creating an evaluator skill while using record. The first is identify high signal source material. The best way to provide feedback to anyone, human or not human, is to bridge the abstract to the concrete. You can't just tell somebody your email writing is trash. You have to come with exact examples. So get a bad example, get a great example, and know why one is bad and why one is great. This third step isn't always obvious. So think through before you press record. The second is you have to turn vibes into checkable criteria. Just saying this looks solid is useless to a skill. So you have to establish a criteria for how to properly validate something. I'm personally a sucker for checklists. So, think through what a checklist for a specific task would be and how to validate if each item on that task list is complete. So, for email, it could be one, make sure the person's name is correct and walk through how you verify it. Two, don't open with dear. Only say hi or hey. Three, use bullets for any action items. You can imagine a checklist of items that will validate if a task is complete. And by doing this while pressing record, it has both a visual and an audio way to think through exactly what you're saying. The third part of creating evaluation skills with record is you have to demand evidence, not verdicts. Require the evaluator to cite the exact reason why something was done. Evidence-based checks are far harder to fake. So have it outline why it made that decision. Building this all up into the email example here. I'll record my screen while I narrate to help build the evaluation. On screen, I'll highlight how I do each of these three steps. When drafting an email, we want to make sure it's a highquality email. So, to validate this, create an email draft skill with an evaluation component. On the left is a bad email. On the right is a good email. On the left, you can see it is a single paragraph that is hard to read visually. On the right, it has clear bullets with concise action items. Before finalizing any draft, make sure to run through our validation checklist. And at the bottom of the draft in red, mark exactly what the results of this checklist were and how it came to that conclusion. I would then stop recording. And I can't stress how powerful this is. And this audiovisisual pairing gives contextual insight that was previously not available. And this record feature isn't specific to Claude. Other tools have it. And in fact, I've used it for some of my clients in a custom tool that I built. Previously, when they would provide feedback on tools that we would build, they would send screenshots and then send us text. But now I created an audio and visual tool that records them using the tool and providing feedback and then I take that to help improve the product. I can't stress how much of a gamecher this has been. Now for years people have been learning to communicate with computers. Now computers are starting to learn how people naturally teach. Most people will use recording to show Claude where they click. But the smart users will use it to show Claude how they think and how they evaluate their outputs. So you now know how to use the record feature to build skills. But once you build it, you have to test it by hand. And at that point, you need to optimize the skill itself, which I do a deep dive on in this video. By taking that and pairing it with the concept I covered here will be a absolute gamecher. So, click the link to that video and I'll see you in the next one. Peace.