Stop Prompting Claude. Show It Instead.

0:00 Claude can now learn how you work just by watching you do it. You do the task, you explain the decisions you're making, and Claude can create the instructions by itself. So, I gave mine a completely

0:12 new version of the task after doing exactly this. And that's when this got interesting because for the last few years, getting better results from AI meant getting better at prompting,

0:23 right? Getting better at explaining what you wanted. But think about it. Prompting rewards people who can explain their process. But a huge amount of what we're good at, we don't know how to

0:34 explain. We know what looks right, what deserves another look, what we would maybe do differently. I mean, imagine you hired someone new tomorrow. You could probably sit down and write a

0:46 30-page manual explaining exactly how you work. Or you could pull up a chair and say, "Come on, sit down and watch me do this." Most of us would probably choose the ladder, right? and Claude can

0:58 now learn that exact same way. You perform the task, you explain what you're thinking as you go, and then Claude can turn that demonstration into a reusable skill. So, I want to find out

1:10 how far can this go. How much of your experience can you actually teach an AI by letting it watch you work? So, I'm going to teach Claude five increasingly difficult parts of how I work. We'll

1:22 start with the process that should be easy and then add preferences and exceptions and actual professional judgment. And for the final test, I'm going to be giving Claude something that

1:31 I never showed it how to handle because remembering what I did is one thing, but figuring out what I meant is something completely different. So, let's do this. Okay. So, before we start, let me show

1:44 you what is actually happening here because this is extremely simple. I click this button that is called record a skill and you start recording. And this way Claude can capture what's

1:54 happening on your screen, your clicks, what you're typing. And if you leave your microphone on, it can also hear what you're saying as well. And you're going to notice that I'm going to talk

2:04 quite a lot while I do this. That sounds like cheating or prompting, but I think this distinction is actually the entire feature because writing down everything you know might be hard sometimes.

2:17 explaining a decision while you're making it, however, is a lot easier. Think about trying to write instructions for how you drive a car. I mean, you would suddenly need to explain a

2:26 thousand things that your brain normally does without even asking for permission. But if somebody sat next to you, you could say, "Oh, I'm slowing down here because I can't see around that corner."

2:37 That's a lot easier. And that is what we're going to do with Claude. We are not sitting down beforehand trying to convert years of experience into a giant mega humongous prompt. We're doing the

2:49 job and then we're explaining the bits that Claude cannot see. I mean, Anthropic actually recommends this because there are decisions that it cannot infer from the screen alone. So,

3:00 we're going to start with something where there should be almost no ambiguity at all. So, we're going to test a simple recurring update, and I have deliberately made it annoying. The

3:11 information comes from four places, and I need to explain to Claude how I do it. I mean, this is the kind of workflow where AI should be brilliant. But writing the instructions for it is also

3:22 the kind of job that you keep putting off because by the time you've documented exactly where everything lives, what to copy, what to ignore, where to paste, what format, you could

3:32 have probably made the report yourself. So instead, I'm going to make it only once. So let's do this together. Record a skill. There we go. Okay. So every week I build

3:45 this same update and it comes from four different places. We have sales first. I don't need every number here. I'm taking the three changes that actually affect this week. So, pipeline value, closed

3:56 one, and the demo bookings. The contacts total and the emails sent don't actually change what anyone does on Monday. So, they don't go in. Now, support. I'm ignoring the one-off complaints. Someone

4:08 not liking where the export button sits is not a pattern. What I do want is this one, the duplicate summary emails, because it's the third week running. And this one, this clinic being blocked

4:20 because that's stopping a customer getting started. Project tracker. I only want what changed status since last week. So bidding has gone green to amber. Onboarding has gone amber to

4:31 green. Everything else held. So everything else stays out. Meeting notes. I'm pulling decisions and open questions. I don't need a record of who said what. So the two headline decisions

4:41 and the two things nobody resolved. And then those go into the same four sections every week. What happened? What's on track? What's at risk? And what needs a decision? And done. And now

4:54 Claude is reviewing what I just did. And look at what it did. I think this is worth pausing on for one second because I didn't write this. Claude has taken what it watched me do and turned it into

5:05 a set of instructions by itself. And now we are going to give it four completely new sources from the following week. And I'm gonna ask it to do the exact same job but with different information. So

5:17 let's see what happens. Okay, I think that's very close. It pulled the right information from all four places. It put it into the right sections and the formatting is basically

5:28 what I did. Of course, there are a couple of things that I would tweak, but for one demonstration, I think that's pretty impressive. So clearly Claude can learn what I do, but this was the easy

5:39 one. There was a standard output. There were standard sources and apart from deciding which details mattered, there was a fairly obvious way to complete the task. A lot of professional work isn't

5:50 like that. Let's be honest. Sometimes two people can look at the exact same material and produce two completely different outputs and they can both be technically correct. The difference is

6:00 preference. So, let's move to test number two. So, I've got a long update here. It's about this project. It has several pages. There are numbers and decisions and progress notes and

6:13 problems, context, and way too much detail for the person that I should be sending it to. My job is to turn this into the short version that I would actually send. And notice what makes

6:25 this different. There isn't one correct summary hiding inside this document, right? I have to choose what is important, what is interesting but maybe irrelevant, what should go first, maybe

6:36 what gets one sentence and what gets completely ignored and disappears from the report summary. So I am going to record myself doing exactly that and let's see if Claude can understand the

6:48 distinctions. So same step I'll click record a skill right this is four pages and I need the version that I would actually send the background section goes every one of them already knows why

7:01 this program exists and it does not change what they do next now data migration 412,000 records 38 discrepancies finished 9 days early that's the biggest number on the

7:13 page and it's the least important thing here because it's finished done things don't need decisions one line at This is the one. Training hasn't started and it's week 11 against a plan that

7:24 said week six. That goes to the top because if they only read one sentence, I think it's this one. Budget 58% spent at 55% elapsed. Contingency untouched. That's fine. So, it's one line and it's

7:39 one line specifically so nobody asks about it. The contractor's leaving in October that needs to stay. It's a small paragraph in a long document and it's the thing that can break December

7:50 stakeholder session 71% attendance that's activity no outcome it goes team and morale can go vendor performance is acceptable so it goes nothing acceptable needs a paragraph and the two decisions

8:03 stay obviously because that's what I'm asking them from okay we are done this is one of those jobs where the final result looks obvious after somebody good has done it but the hard part was not

8:15 writing the six sentences. The hard part was deciding which six sentences were worth actually keeping because summarizing isn't really about making something shorter. It's deciding what

8:25 deserves to survive and explaining in a prompt how to do that. That's the hard part. But now Claude has watched us make those decisions. Okay, so let's give it a completely different long update and

8:37 see if it is able to apply the same logic. And before Claw does anything, I'm actually going to create my own short version separately. Done. Now claw. Okay, let's run the

8:55 skill. Okay, let's compare. So I kept this item and it kept that too and I dropped this other one. Okay, interesting. it led with this other one whereas I led with

9:10 something else. And this is useful because now we're getting away from copying a workflow. We're asking whether a demonstration can transfer a preference. And I think this is

9:20 surprisingly close to the version that I made. It's not identical, which I wouldn't have wanted it to be, but it does seem to have captured what I tend to value. But now the next test is where

9:31 I'm expecting this to become a lot harder because choosing what I keep is still visible. I can highlight a paragraph. I can delete a sentence. Claude can see the consequence of my

9:41 choice. But experienced people make another kind of decision all day. They notice something that does not look right. Sometimes before they can even explain why. I'm sure you've been there.

9:51 By the way, I'm putting all five skills from this experiment, plus the exact narration prompts that I use while recording them, in a free downloadable inside our founders community. The

10:02 community is paid, but you have 7 days for free. So you can go and grab that if you don't want to stay longer, although I think you should. So if you want to try this yourself, you don't need to

10:11 reconstruct anything that I'm doing in the video. You can watch closely and then go and practice on your own. So let's look at these numbers. These are, of course, fictional business numbers

10:21 that I've created for this experiment. And I've designed them to be slightly annoying because some of the biggest movements here are completely normal. And one of the smallest changes is the

10:33 one that I actually want Claude to worry about. I know I'm a little mean. This is much closer to what experience feels like. Actually, you don't look at every movement and scream, "Oh my god, there's

10:44 a problem. There are no alarm bells." You've seen enough cycles to know some things move and some things spike after a launch and some things fall at the same time every year and some numbers

10:56 are supposed to be boring, but when one of those moves, you pay attention. So, same move. Record a skill. Okay, traffic is down 18%. That's the biggest movement here. And I am not chasing it because

11:10 the paid campaign ended 11 days ago and I would expect this to fall. Support messages up 22%. We shipped a release on Monday. People ask more questions when something changes. I would worry if this

11:21 was still up in 3 weeks with nothing behind it, but for now it's fine. Trial to paid flat at 4.1. It's good. Some numbers are supposed to be boring and this is one of them. Proposal requests

11:33 up 31%. Big move, but this always moves at the end of the month. Watching it, but not escalating anything. Email open rate down 4%. That's noise. It bounces around by that much every week, so I'm

11:47 not worried. Now, refund requests 3 to 7. That is the smallest change on this page in absolute terms, and it's the only one that I would actually chase. This number has sat between two and four

12:00 every week for over a year. It doesn't usually move. So when it moves at all, something caused it. And I want to know what before it becomes 20. And that's the thing worth saying. A number that

12:12 actually swings can swing a long way and mean nothing. A number that never moves only has to move a little to mean a lot. All right, let's create the skill and give it a test. And while it does that,

12:25 I want you to notice what's going on here. The data doesn't contain a label saying Laura considers this unusual. That comes from context. It comes from having seen the number before. It comes

12:37 from knowing what usually changes and what usually doesn't. A workflow is what you do when nothing unusual happens. Expertise shows up when something does. And I don't know if you know, but there

12:49 is a name for a lot of this. It's called tacit knowledge. We know how to do things that we've never turned into a written procedure. We just know. We know when an email feels off. We know when a

13:01 customer is more upset than their words might suggest. We know when a project that is technically green is about to become red. And we also know when a number looks wrong. And sometimes if

13:12 somebody asks, "How did you know?" Your first answer might be, "I don't know. I've probably seen this before." I mean it has happened to me so many times over the last 21 years. It has happened back

13:24 in my consulting days where I would see something and associate it with past experience. It happened in our agency. It even happened recently when analyzing the performance of a YouTube video. And

13:37 that experience changed what I pay attention to now. And that's why this test is so important to me because experience is often knowing what to ignore. The beginner sees 10 signals,

13:47 but the experienced person knows which two deserve another look. So, we've shown Claude one set, and now I've created a second set with different movements and different traps, but I

13:58 know which one I think deserves escalation, and Claude doesn't. So, let's run it and see if he got the gist. I think it got it. And more importantly, look at the reason that it gives. It's

14:11 so interesting cuz it didn't choose the largest change. It chose this one because based on what I demonstrated, this is the type of number that should normally be stable. That is a very

14:21 different thing from flag anything that moves more than 10% that you would put in a normal prompt for a skill that you would write. So this is the point where I think the entire idea becomes even

14:32 bigger than recorded skills as a feature. The bottleneck in AI hasn't only been what the model knows. It's been our ability to transfer what we know into the model. because we've spent

14:44 years asking how do we write better prompts. But maybe another useful question is how do we make our judgment visible enough for AI to learn from it? Because we're about to remove the

14:56 numbers. Okay, there won't be an anomaly to spot. There won't be a correct answer. There are just 20 things competing for five slots in my hypothetical week. Let's see if it's

15:07 then able to make decisions the way that I do. Okay? Because this is a very normal professional problem. You have 20 things that you could do this week. You have enough time and attention to do

15:17 maybe five of them properly. And everything else needs to get delegated, delayed or dropped. But then some are urgent and some are important and some look important because somebody senior

15:28 asked for them. And some take 10 minutes and then others take 3 hours or days. But also a couple might unblock other people. And maybe one has been moved from week to week. so many times that at

15:41 some point you might need to admit that you're not going to do it. I know we all have one of those, but we're not going to talk about mine today. Okay, so record a skill. Number one, the pricing

15:54 page copy. Taking it first. It's not the biggest task on here, but three people are waiting on my decision, so 20 minutes for me gives them the rest of the week back. Number two, the board

16:04 summary. It says Friday, but nothing actually breaks if it lands on Monday, so it can move. Number eight, the duplicate email bug. Customers have raised this four weeks running and

16:13 nobody has picked it up. So that needs to be in. Number 11, the onetoone with the newest person. I mean, they've asked for it twice. That needs to be in. And honestly, it should have been in 2 weeks

16:24 ago. Number 10, the vendor demo. I could take it. That's not the same as me being the right person to take it. So, this one's dedicated. Number 14, the partner's question about headcount. This

16:36 came from the most senior person on the list and it's still going to the delay pile because it was asked casually and there is no decision hanging on it this week. Number five, reorganizing the

16:47 sheer drive 11 weeks. I'm going to be honest and admit I'm not going to do this one. Number 18, rewriting the sales one pager. It's uncomfortable. It needs actual thinking and that is exactly why

16:59 I have moved it 3 weeks running so it goes first not last. And notice there isn't one rule doing it. Sometimes it's who's blocked. Sometimes it's what breaks if it waits. And sometimes it's

17:11 whether somebody else could do it. And also sometimes it's me noticing that I've been avoiding something. Because prioritization is where your values become visible under pressure. That's

17:20 why your calendar and your task will probably tell you more about how you work than almost any productivity philosophy that you could write down. Okay. So, my skill is ready to be

17:32 prepared. These are my five. Now, Claude has a skill built from that demonstration, and I've created a 20 completely new tasks list. Before Claude sees them, I'm making my own five. And

17:47 then we can compare. Done. Let's run it. Okay. So, this one is staying. This one's delegated.

18:01 Interesting. So, let's compare. I chose this one and so did Claude. And the reason is basically the same. It also delegated this one, which I did, too. But this one's interesting. Claude kept

18:16 it, but I dropped it. And we had different reasoning. But I I see where it's coming from. I think this is much closer than I expected. Honestly, it's not making my exact choices every time,

18:27 but the choices rhyme with mine, so to speak. And I think that's the more interesting result because I don't need clot to become a photocopy of me. And probably you don't want that either. You

18:36 want to know whether some of the judgment that you've accumulated can become useful context for the next decision. And that brings me to something that I kept coming back to

18:46 throughout this experiment. I think for the first time, our years of experience might make AI easier to use and not harder. Because think about the last few years. If somebody younger than you knew

18:58 how to write better prompts, understood every new AI tool, and spent every evening playing with automations and agents, they had an obvious advantage. But you have something else. You know

19:09 the work. You know what matters. You know what usually goes wrong. And you know which problem deserves attention before anybody else notices it. And now we are getting interfaces that let us

19:20 transfer more of that experience without first becoming professional instruction writers. Right? Experience is compressed judgment and recorded skills give us a new way to unpack some of it. But I

19:33 haven't tested the hardest part yet because Claude just learned how I prioritize my week. Maybe that's all that it learned. So, I'm going to take myself out of the equation. And the next

19:44 test is to see if judgment can transfer. And I want you to meet Mark. Mark is fictional, but I've built him to look like somebody who could easily be watching this video. Let's assume he's a

19:57 sales director. He manages a team of 12. His three priorities this quarter are protect the renewals coming up, help two new team members become productive, and get the team ready for a major product

20:09 launch next month. Seems simple, right? He however has different responsibilities than me. Different goals, different constraints, and also a very different week. So now we're going

20:21 to take his list of 20 tasks and I'm going to give Claude Mark's role, his goals, his constraints, and these 20 things competing for his attention. What I'm not giving Claude is a new framework

20:35 for choosing the five. That's the test. If this skill only learned which tasks does Lara like, I think it will struggle. But if it learned something closer to how does Lara reason about

20:48 limited time or consequences delegation and who gets unblocked, then some of that reasoning might transfer. Because if a principle only works on the person who taught it, it may not be a principle

21:00 yet. And before we run this, I'm going to make this slightly more dangerous for me. I'm going to write down where I think Claude will get it wrong. So prediction number one is that I think

21:13 it will overvalue this one and I also think it might also undervalue this one because the payoff isn't immediate. And prediction three is that I think it will automatically keep

21:26 this one because of who asked even though Mark's actual goals point somewhere else. I'm writing these down so I can pretend afterwards that I saw the result coming. Okay, now I'm going

21:37 to choose MarkX 5 myself. And I'm not choosing what I would personally enjoy doing. I'm looking at his goals and his team and what only he can do, what can be delegated, and what creates

21:49 consequences if it waits. Okay, done. Now it's Claude's turn. We're going to run the skill and see what it does. All right, let's put them side by side. So, first one looks like we agree. Second,

22:06 we agree again. Third. Okay, so this is the one that I did predict. I thought that Claude would go for this one and it actually did. That's interesting. I was wrong because the reasoning it gives is

22:21 not what I expected it would do. That is not a rule that I explicitly gave it for Mark. So, it is applying something from my demonstration to his constraints. I'm impressed. So final score, Claude

22:34 matched me pretty pretty well, but the raw number is almost the least important part. Look at where it matched and look at where it didn't. Okay, if it got something right because the new

22:46 situation looked like the old one, that's pattern matching. But if it got something right even though the surface situation changed and the reason still makes sense for mark goals, that's much

22:58 closer to the thing that I was looking for. Now look, at the beginning I said the interesting question wasn't whether Claude could remember what I did. It was whether Claude could figure out what I

23:07 meant. And I think we have an answer now. So can Claude learn how you work by watching you? Clearly yes. But that sentence needs an asterisk. The mechanical stuff is the easiest. When

23:20 the inputs are clear and the outputs is repeatable, one good demonstration can give Clot a surprisingly useful starting point. Preferences are harder because your choices need to be visible.

23:31 Exceptions are harder again because the important information may be the context that you carry in your head. And then judgment is where it becomes really interesting because a demonstration can

23:40 expose pieces of how you think that you might never have bothered writing down, but it doesn't read your mind. Right? If a distinction matters and Claude can't see it, you still need to say it. And

23:53 that is why I don't think the lesson from this is prompts are dead. Clearly they're not. The lesson that I would draw from this is that we have another bandwidth for teaching AI now because

24:05 before we mostly transferred a workflow through written instructions. Now we can transfer some of it through behavior. We click here, we skip that, we choose this, we ignore that. And while we're

24:17 doing it, if we explain the tiny decisions that normally disappear inside the experience, we can get it even closer to what we expect. Prompting asks you to describe your work from memory,

24:29 but demonstration lets you explain it while the work is happening. That difference sounds small until you try to document something that you've been doing for 15 years. Trust me, then it

24:39 becomes enormous. And I think this is the bigger shift. For years, we've talked about what AI knows, how smart the model is, how capable it is, how much context it has or can hold, which

24:50 benchmark it beats. But honestly, for professionals, there has always been another bottleneck. How do I get what I know into the AI? Because the valuable part of your work is rarely the obvious

25:03 sequence of clicks, right? It's knowing which sentence stays and which number looks wrong. which task deserves Thursday morning and which urgent request can wait until Monday. I mean,

25:16 there's all of that reasoning that is not really easy to put into a prompt. The clicks are visible, but the judgment is the job. And if AI gets better at learning that judgment from

25:27 demonstrations, the people with experience have something extremely valuable to teach it. That is the part that I think is easy to miss. The AI conversation has made professional

25:37 experience occasionally feel like baggage. You hear about a new model, a new workflow, a new agent every week, and the person who spent 6 months experimenting with all of them can look

25:47 more capable than the person who spent 20 years learning the actual work. But those are two different forms of knowledge. One is knowing how to use the machine. The other is having something

25:58 worth teaching it. And I think for the first time, your 20 years of experience might make AI easier to use than harder. Not because Claude suddenly absorbed 20 years of expertise from a 10-minute

26:08 recording. No, it didn't clearly. But we just watched the limits. The interface between your experience and the model just got easier. And I think that's the useful difference or change. So if you

26:22 have this feature, I would not start by asking what amazing skill should I build. Look at your actual week. Find something that you do repeatedly that maybe you have never properly

26:33 documented. Maybe you build the same update every Friday. Maybe you turn a long document into a short executive summary every week or every month. Maybe you review numbers and know which ones

26:44 deserves attention. Or maybe you make a decision 20 times a week that somebody new on your team would need maybe 6 months to learn. So record yourself doing one. And while you're doing it,

26:56 don't narrate the clicks. Claude can see the clicks. What you need is to narrate the judgment, the way you make decisions. So instead of now I'm opening this file, you can say I'm opening this

27:09 file because this is the source that I trust when these two numbers disagree. And there is one bigger consequence of all this. Teaching Claude one job well does not mean Claude understands the

27:19 rest of your work. We just spent this entire experiment building context around individual decisions. Your tools, your preferences, your recurring work, the information that Claude can access.

27:30 All of that still determines what happens when you ask it to do something new. So, I said at the beginning that Claude can now learn how you work. Well, we've now watched it learn a process,

27:40 pick up preferences, attempt exceptions, imitate judgment, and then try to carry that judgment into somebody else's week. It isn't mind readading and it's not perfect, but show me once has become a

27:52 legitimate way to start teaching AI how you work. And I think that's a much bigger deal than saving 10 minutes writing a prompt. So, if you want to try the exact five skills from this video,

28:02 plus the narration sheet showing what I said while I recorded each one, I have put those in a free document that you can download inside our founders community. The link is going to be down

28:12 below as well as here. And then you can also watch the complete Claude setup because once you've taught Claude one job, well, the next problem is making sure the rest of Claude is actually set

28:22 up around the way you work. So, you can go ahead and watch this video here. In the meantime, thank you so so much for watching. Like this video if you did. Be sure to subscribe if you haven't done

28:30 so. Share this with anyone in your circle of friends or family or co-workers who has a lot of experience to share with Claude but finds it hard to do it through prompts. Until next

28:39 time, I hope to see you in our communities. Whether you want to join our trailblazers or our founders hive, there's loads of value there and thousands of people who are waiting for

28:48 you to share the same path that I'm sure you're on as well. So, I hope to see you there. Fight.