Learning Library

← Back to Library

No‑Code Digital Twin Prompt Walkthrough

Key Points

  • After publishing a long, technical guide on building digital twins, the author received requests for a simple, no‑code solution that everyday users could apply without an enterprise setup.
  • To meet this demand, he created a single “system‑level” prompt (named V2) that walks a user through setting up a digital‑twin simulation step‑by‑step, defining the AI’s role, mission, and workflow in one cohesive script.
  • He demonstrates the prompt with two realistic conversation simulations—a salary‑negotiation role‑play and a product‑approval office‑politics dialogue—to show how the digital twin can emulate human interactions.
  • The video compares how the prompt performs under GPT‑4.0 versus GPT‑3.5, highlighting clear practical differences in output quality and responsiveness between the two models.
  • The entire experiment is run in the Perplexity web‑browser AI assistant, which the author finds useful for parsing complex prompts and monitoring the simulation’s progress.

Sections

Full Transcript

# No‑Code Digital Twin Prompt Walkthrough **Source:** [https://www.youtube.com/watch?v=pebgrFQ-C7M](https://www.youtube.com/watch?v=pebgrFQ-C7M) **Duration:** 00:26:42 ## Summary - After publishing a long, technical guide on building digital twins, the author received requests for a simple, no‑code solution that everyday users could apply without an enterprise setup. - To meet this demand, he created a single “system‑level” prompt (named V2) that walks a user through setting up a digital‑twin simulation step‑by‑step, defining the AI’s role, mission, and workflow in one cohesive script. - He demonstrates the prompt with two realistic conversation simulations—a salary‑negotiation role‑play and a product‑approval office‑politics dialogue—to show how the digital twin can emulate human interactions. - The video compares how the prompt performs under GPT‑4.0 versus GPT‑3.5, highlighting clear practical differences in output quality and responsiveness between the two models. - The entire experiment is run in the Perplexity web‑browser AI assistant, which the author finds useful for parsing complex prompts and monitoring the simulation’s progress. ## Sections - [00:00:00](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=0s) **Simplified Digital Twin Prompt Demo** - The speaker introduces a streamlined, code‑free prompt for building personal digital twin simulations and showcases its use through salary‑negotiation and product‑approval dialogues while comparing model behaviors. - [00:03:17](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=197s) **Building Runnable Simulation Prompts** - The speaker outlines a step-by-step process for confirming data, embedding inline tables, defining output rules, and transitioning to a digital‑twin simulation prompt that reliably moves from information gathering to execution. - [00:06:26](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=386s) **Ensuring Accurate World‑Building via Confirmation Phrases** - The speaker explains that a confirmation phrase forces the model to actively listen, summarize, and verify user inputs before filling a digital‑twin prompt template, guaranteeing consistency and allowing customizable purpose, mode, and effort settings. - [00:10:09](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=609s) **Explicit Prompt Contracts for Runnable AI** - The speaker outlines how a meticulously defined, progressive‑disclosure prompt—detailing purpose, mode, effort, instructions, references, and output—forms an explicit contract that forces the model to start immediately within a single runnable block, solving the data‑embedding challenges of the V2 system. - [00:13:47](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=827s) **Designing Robust Prompt Scripts** - It explains how framing responses with a confirmation template, specifying a runnable prompt contract, and using a fixed question order creates a deterministic, state‑machine‑like workflow that improves LLM reliability while stressing progressive disclosure and careful placeholder handling. - [00:17:31](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=1051s) **Creating Corporate Stakeholder Personas** - The speaker explains how they fabricated realistic executive characters—CTO, director of engineering, CFO, and CEO—with specific worries, highlighted open legal and pricing challenges, and were instructed to generate a 500‑word scripted transcript for a mock debrief. - [00:22:00](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=1320s) **AI-Driven Negotiation Role‑Play Review** - The speaker demonstrates how an AI tool simulates a CPO‑finance negotiation, provides feedback, and compares outputs across different language models. - [00:25:47](https://www.youtube.com/watch?v=pebgrFQ-C7M&t=1547s) **Super Prompt for AI Digital Twins** - The speaker outlines how a single, adaptable ChatGPT prompt can construct versatile digital twins for negotiations, interviews, and other multi‑party scenarios, demystifying the process and highlighting its broad applicability. ## Full Transcript
0:00You know, a few days ago I did an entire 0:02piece on setting up digital twins. It 0:04was very large. I think I wrote a 0:06hundred some page guide for it. The 0:08point is people took it and said, "This 0:10looks really hard. You wrote all these 0:12pages for it. Help me figure out how to 0:15do it simply. I don't have an enterprise 0:17setup. I'm not making robots learn to 0:19walk in a warehouse. This is not a 0:21Fortune 500 company. I want to use 0:24digital twins without having to write 0:26code." And I took that as a personal 0:29challenge. And so what we are going to 0:31work through today is an actual prompt 0:34to set up your own digital twin 0:38simulation. We have a couple of goals as 0:40we work through this prompt. We want to 0:41understand first how the prompt 0:43functions. It's another one of those 0:45system level prompts. So it actually 0:46walks through an entire process or flow. 0:50It creates a scenario and then it walks 0:53you directly through it all in one 0:54prompt. And we'll see how it works. The 0:56second thing I want to do is I want to 0:58show you some of how it works with 0:59actual real conversations that I had 1:02using that prompt and we'll walk through 1:04them. You'll get both a salary 1:07negotiation conversation because 1:08sometimes we want to game those in 1:10advance and you'll also get a classic 1:12like product approval conversation to 1:14give you a sense of office politics and 1:16how it simulates office politics. It's 1:19interesting to see how different models 1:22affect this simulation. And for just one 1:25more wrinkle, I'm going to give you a 1:27sense of how chat GPT40 1:31compares to chat GPT 03 handling this 1:34prompt and running the simulation. This 1:36is actually one of the clearest examples 1:38I've seen of the practical realworld 1:42differences between those two models. So 1:45with that, let's get to the prompt. 1:47Okay, here we have the digital twin 1:49stakeholder prompt that I constructed. I 1:51helpfully named it V2 because it took 1:54multiple iterations to get this prompt 1:56working and I want to walk through how 1:58it functions. As before, I am running 2:00this in the web browser comet with the 2:03handy assistant pulled up. I had a few 2:05questions when I did this last time, so 2:06just so you know, this is the Perplexity 2:09web browser and that's an AI assistant 2:11that Perplexity launches with the 2:13browser. I find it useful. I did a write 2:15up on it. Uh, and I find it really 2:16helpful in situations like this where 2:18I'm trying to understand what is going 2:20on with a complicated page. So, as you 2:22can see, the prompt as a whole is quite 2:24large. I will just scroll to the end of 2:26the prompt and you'll sort of start to 2:27see, wow, there's a lot here. So, let's 2:29get into it piece by piece. First, we 2:33set the role. This is not surprising. I 2:35think I've talked before about the idea 2:36that you're setting the role as a way of 2:38invoking a particular semantic space. 2:40And next, you define the mission. In 2:42this case, we have a four-part mission. 2:43This is part of why you'll see better 2:46performance out of 03 than 40 when we 2:48get later into this video. Number one, 2:51get the information out of me to figure 2:53out what what is needed for a realistic 2:56multistakeholder negotiation in my 2:58situation because this is not a specific 3:00prompt for compensation negotiation. It 3:03is a super prompt that actually helps 3:05you to set up whatever simulation you 3:07want to run. Second, ask only one 3:09question at a time. I find it so 3:11overwhelming when LLM's ask too many 3:13questions. So I I tend to include that. 3:15Third, confirm the answer. And then 3:17fourth, when the answers are gathered. 3:19And now we get into the heart of it. You 3:21need to generate a runnable simulation 3:24prompt that embeds every confirmed 3:26detail, contains inline data tables if 3:29the user has not provided external 3:30files, but you can, and includes clear 3:33output rules, and a begin delimter. And 3:35then five, 3:37write the prompt out as plain text and 3:39immediately transition into simulation 3:41mode. One of this was actually one of 3:43the reasons that I went to V2 is because 3:45I found when I was playing with this 3:46prompt and working with it, it is 3:48difficult to get the AI to reliably go 3:51from I'm learning about you and 3:53gathering information to okay, we've got 3:57the information. Now, it's time to 3:59actually run the simulation and be a 4:01digital twin. In this case, multiple 4:02digital twins. So far so good. And 4:05what's interesting is when you get to 4:08the the end of the prompt, it 4:11immediately invokes 4:14it immediately invokes the twins. And so 4:17we'll get down here and see what this 4:19means. But I want you to see the 4:20connection between five where we talk 4:22about what we want the twins to be and 4:24the end here where we talk about each 4:26twin's opening statement. And that's the 4:27way you begin. because this ties this 4:30ties the tokens back when the LLM is 4:34reading this prompt so that it knows 4:36okay we have to remain in character 4:38opening statements of every twin I see 4:40that back here it's like we're doing 4:42this memory management across a larger 4:44prompt and this is a great example of 4:46how that can work okay now the LLM at 4:49this point would still not have all that 4:51it needs to run so let's get to the 4:53question script because we need 4:55something that gets information out of 4:57you one at a time. The obvious first 5:00one, what situation are we negotiating? 5:02List the participants. None of these are 5:04super surprising. List the success 5:06metrics or how you characterize a win. 5:08What twin will you play? And by the way, 5:11this works even if you don't want to 5:13play a twin. So, if you just want to see 5:14the LLM game it out and read the script, 5:17you can say, "I won't play any of the 5:19twins. You're playing them all." And it 5:20will. How many turns before timeout? So, 5:23this is framed as a multi-turn 5:24conversation. And so you can say three, 5:26you can say six, you can say whatever 5:27number you want. Please provide key 5:29numbers or attach a file for deals or 5:32comp. Otherwise, give ballpark figures. 5:34If you're salary negotiating, it's 5:36obviously comp. But when I was doing it 5:38for my uh product sample that I'll show 5:40you in a moment, I ignored the comp part 5:42and it said deal. So I put in like deal 5:44data and like where our pipeline was at 5:46and all of that. These are all madeup 5:47numbers. Um, and I felt really free to 5:50add the data I wanted. I will also call 5:53out that this is a somewhat flexible 5:56user response prompt. It is designed to 5:59be a place where you can put a lot of 6:01information and you'll see in one of the 6:03sample chats how I'm able to pack in 6:06like 300 words of information which is 6:08really more than I designed the prompt 6:09to take and the prompt is able to handle 6:11that. Okay, constraints and policies 6:13that's always important in any 6:15negotiation and then your output 6:17preferences, transcript style, word 6:19limit, etc. So at that point uh I 6:22actually define a confirmation phrase. 6:24Now why do I define a confirmation 6:26phrase? Why do I say after the user 6:29answers each question, this is what you 6:31should say? Well, number one, I define a 6:34confirmation phrase because it is 6:37critical to actually understand what the 6:40user said to get this world building 6:42right. With a lot of prompts, if you 6:44write the prompt, you can do a second 6:45turn and kind of figure it out if you're 6:47trying to refine. But with a world 6:49building or digital twin kind of prompt, 6:52the world has to be right from the 6:54get-go. And so I need to make sure that 6:56it understands, right? And so I have it 6:59come back and actively listen and 7:01summarize on purpose. Okay, so it's 7:04actively listened. It's collected all 7:05the information. We now get to the 7:08runnable prompt template. When all the 7:10answers are confirmed, please fill the 7:12placeholders and output you are the 7:14digital twin negotiation arena host. So, 7:16in a sense, you might be wondering, why 7:18the heck is this here? Purpose, mode, 7:20instructions, reference. Haven't we been 7:22talking long enough? Why does this go 7:25on, Nate? Well, I'll tell you why it 7:27goes on. It goes on because we are 7:30trying to bake in some degree of 7:33consistency in this prompt and giving it 7:36a purpose, a mode, an effort level, 7:39which by the way, you can set 7:40differently. Like I set it at high, but 7:42if you want to change this prop before 7:44running and set it at low, you can. Uh 7:46you can also set the scenario and 7:50hardcode it here if you want or it will 7:52fill in the scenario for you based on 7:54your answers. It will then output all 7:56the rest of this based on your answers 7:58and the reference. And as it does this, 8:00it is literally encoding into the stream 8:03of conversation everything it needs to 8:05know to keep it going. This little piece 8:08here does the important job of acting as 8:12a conversational anchor. It acts as a 8:15conversational anchor so that the rest 8:17of the world building will work. Right? 8:19If you have ever tried to do digital 8:21twin stuff, you know that the 8:22personality is drifting can be a 8:24problem. This is part of how we control 8:26for that. Having this clear reiteration 8:29before we start. Finally, we begin 8:31insert each twin's opening statement 8:33right after the delimiter. So that's the 8:35prompt. That is what we will run. That 8:37is what I will show you. Before we go 8:39further, I want to actually go down. I 8:41actually broke this out into principles 8:43that we can look at. Uh, and I also 8:46broke out takeaways. And so I want to 8:48spend a second 8:50looking at the takeaways and the 8:51principles and make sure you understand 8:53why I did what I did before I go to the 8:56actual examples. So principle number 8:57one, identity lock in. I want to make 8:59sure that it's a digital twin and I want 9:01to make sure I invoke that corner of 9:02latent space. It is a deterministic 9:05state machine which is a fancy way of 9:07saying I am deliberately creating a 9:09fixednumbered question script to gather 9:12input on a path. This turns an 9:14open-ended chat into something that is 9:16very repeatable and something where the 9:19uh process can be deliberately shifted 9:22for ease of use in learning about 9:24different future timelines. So, if you 9:26want to game out a timeline where you 9:29open with 210 for your comp or a 9:31timeline where you open with 180 or a 9:33timeline where you open with 150, you 9:35can do all of those things in separate 9:38chats and you get a really tight 9:40controlled uh simulation of how that 9:43conversation might unfold. Similarly, if 9:46you're doing a job interview, which you 9:48could use for this, you can game out in 9:50multiple different chats, what if I 9:53answered this way? and you can watch 9:55your other stakeholders respond and 9:57actually like game that through and 9:59think that through more easily. Humans 10:01are not great at simulating entire 10:03digital scenarios in our head with 10:05multiple stakeholders at high fidelity. 10:08That is what this prompt is designed to 10:09do. Progressive disclosure. We talked 10:12about asking one question at a time. 10:14Echo confirmation loop. Rephrase. 10:16Actively listen. Actively listen might 10:18have been a better way to put this. 10:20Explicit output contract. This is the 10:23contract that we described up here that 10:26sets the terms of the debate. It says 10:29here's your purpose, here's your mode, 10:31here's your effort, here's your 10:32instructions, here's your references, 10:34and here's your output. It's really 10:36important to be clear and explicit about 10:37that when you're doing digital twin 10:39work. Okay. And then we get to handoff 10:42embed all the requisite data inside the 10:44runnable block. This was one of the 10:47challenges I had and this is why we 10:48named this V2. It is hard to get all the 10:52data inside a runnable block, collect 10:55all of it, make it run inside the same 10:58prompt. One of the keys to getting it to 11:00do that is to be very very explicit 11:02about that reiteration of the contract 11:04and explicit in the questions about what 11:07you are gathering. And so if you look 11:10back here, we are being very explicit 11:12about what we want in this question 11:15setup here, these nine questions. 11:18And we are specifying it needs to be 11:20runnable and we are specifying it needs 11:22to begin now. We are not giving the 11:24model any choice. We are saying here's 11:26all the information and you must begin 11:28now. And that was very deliberate. 11:30Context switching invokes the model's 11:32opening moves very deliberately and we 11:34define what we expect for the first move 11:36which helps the model get into that 11:38space. We say right up here, we need you 11:41to make opening statements that because 11:43you could begin a lot of different ways, 11:44but we say opening statements because 11:47that enables the model to enter the 11:49simulation space in a predictable way. 11:52If you leave that blank, you are giving 11:55the model cart blanch to open any way it 11:58wants. And you don't want that for a 12:00predictable simulation builder. You 12:02actually want a little bit of 12:03predictability so you can start to 12:05insert variables and learn. Okay. From 12:08there we get into uh visual parsing and 12:11bounded variables. So users can set 12:13rounds, they can set word limits, users 12:15can um actually see what is going on 12:17easily because we're using those asy 12:19gutters like the delimiters. I don't 12:21want to say these are fantastic 12:22principles. Like don't make me tell you 12:25that begin now with special asy gutters 12:27is somehow going to be more magical than 12:29begin now. It's not. But it sure is 12:31easier to read. And if we're reading 12:33these large prompts like I've been 12:34describing, it is sometimes nice to have 12:36clean gutters and bulleted lists. 12:38Finally, there are error rules for 12:42failure modes. And part of the error 12:44rules I've already called out to you. 12:45Like if you go back up here, one of the 12:48hidden error uh rules is the 12:51confirmation phrase. If there is a 12:53problem, you are going to see it because 12:55it's going to come back and tell you 12:57because the user answer paraphrase is 13:00going to be correct. But there's other 13:02ones too like do not ask further setup 13:05questions after the honorable prompt is 13:06admitted. We are scoping down so it 13:08doesn't go forever. We are demanding 13:10that it remain in character. We want to 13:12ensure it retains the detail. We are 13:14demanding that it only asks one question 13:16at a time. We are giving it a lot of 13:17constraints. Okay. Now I want to finish 13:20by talking about the structure and 13:22starting with the system ro declaration 13:25is classic. Getting to the mission is 13:27the correct overall next step. And by 13:29the way, these are larger pro uh prompt 13:32structures that you can use in other 13:34prompts, not just for these super macro 13:36prompts, right? If you start with, hey, 13:38here's my system role to move you into 13:39latent space. Here's the mission that I 13:41have for you. Uh and then here's the 13:43content that I want you to engage with. 13:45In this case, it's the script. you're 13:47going to be in a good spot for a lot of 13:49different prompts. If you frame how you 13:52want the response to work, which is the 13:53confirmation phrase template in my case, 13:55but could be something totally different 13:56for a prompt of yours, it's going to be 13:59really, really helpful. And finally, if 14:01you specify a contract and how you 14:04begin, if you're trying to do a super 14:05prompt like this and you specify a 14:07runnable prompt contract at the end and 14:09you also specify how you want the model 14:12to begin, it increases the odds that 14:14your prompt is going to actually run 14:16successfully. Okay, we are going to skip 14:19the micro details that add delight. 14:20We've talked about the asy. We've talked 14:22about markdown. Aren't you glad I can 14:24make things pretty? Uh we will talk 14:26briefly about how these pieces reinforce 14:28one another. The fixed question order 14:30reinforces the deterministic state 14:32machine. The builder never deviates from 14:34the script and always gets the questions 14:36the same way. And you can treat it like 14:38an automaton in that regard. The only 14:40difference is the power of the LLM 14:42behind it, which is really interesting 14:45because you'll see in the examples how 14:4803 versus 40 is different. If you are 14:51teaching this, and I know some folks 14:53that watch my videos teach my work to 14:55others, and that's fantastic. You want 14:57to remind people of the principle of 15:00progressive disclosure and the 15:02importance of template integrity. Be 15:03careful with your placeholders. Be 15:05careful with asking too much of the 15:08model in one go without those checks and 15:10balances that I showed you in the 15:12prompt, without asking it to reiterate 15:13the contract, without asking it to be 15:16specific in its summary as it comes back 15:19to you. We are giving the prompt the 15:22token scaffolding it needs to be 15:24successful. Finally, potential pitfalls. 15:27You want to make sure that you are 15:30giving it limits. And so that is one of 15:32the things that we did not go as 15:35explicit here that I find that the model 15:37tends to get to in practice, as we'll 15:39see. You want to not be 15:42overcommitting the model to responses it 15:44can't deliver. Another example that 15:46would have strengthened this prompt more 15:47that you might want to use if you add 15:49more files is to summarize key numbers 15:52in two to three lines to force the model 15:53to answer back on files because you 15:56don't want it to just skip over the file 15:58content. So there are little things that 16:00think of these is essentially tweaks you 16:02can make depending on what you're 16:03looking to accomplish. We've gone on 16:05long enough. Let's move to digital twin 16:09negotiation builder v2 using 03 and see 16:13how the conversation went. Here we are. 16:16I run the prompt just as you described. 16:18The situation we're negotiating is a 16:21product pitch LLM to SQL project. It 16:24understands it. It comes back. Here's 16:25who I'm pitching to. CEO, CTO, CFO, 16:28director, me and the CRO 16:31here. I I want approval to build the 16:33product and launch. That's my goal. I 16:34have been in these situations. I'm 16:35picking something I know well so I can 16:37assess the quality. Next question. Which 16:39twin will you play? I'm going to play me 16:41and be the director of product. Great. 16:43How many turns before timeout? Let's do 16:45three rounds cuz most CEOs don't 16:47tolerate more than that. And then I 16:48throw it a curveball. This was not 16:50planned for in the prompt. I throw it 16:53300 words on my product and I say this 16:56is what I pitched them. Right? It's all 16:58about how amazing this imaginary product 17:00is. I I am okay with the numbers being 17:02made up because this is an example and 17:04this is my ask Q3 roll out request. Got 17:07it. Next question. Provide any key 17:09numbers and I basically say I could have 17:12said just stick with these numbers up 17:13here but I want to think about the deals 17:15in this case because I think that's 17:16going to matter for the simulation. Most 17:19most seuite conversations in a B2B SAS 17:22have to do with deals at some level. I 17:24say we're chasing million-dollar deals 17:25with enterprise clients. We've had a 17:27pipeline slowdown in the last 6 months. 17:29Okay, great. The key data we're pursuing 17:31is this. It reiterates. It asks me to 17:33talk about the different characters. And 17:34so this is me like you basically make up 17:37the characters you want. If you don't 17:39know, you have to infer and guess. I had 17:41fun making these up. These are 17:43fictional, but they're plausible, right? 17:45The CTO is worried about hitting the 17:47schedule. The director of engineering is 17:48somewhat hostile because he thinks we're 17:49we're missing on core product. The CFO, 17:52she's worried about close rates and 17:53pipeline. The CEO, he's a business guy. 17:56He believes in enterprise and he loves 17:57AI, but he doesn't know how it works. 17:59Like these are all very plausible and I 18:01need a promotion, right? Like because I 18:02I' I've definitely been in that spot. So 18:04there it is. And then it asks, "What are 18:07the open legal or approval issues?" 18:09Super simple. Uh there's an open legal 18:11issue around LLM and data privacy. 18:13There's also an open question around 18:14pricing for usage, which is a huge issue 18:17more broadly. Different video. It 18:18reiterates and I say, "Please generate a 18:21script style transcript. Set the limit 18:22for debrief at 500 words. Leave room to 18:25speak." Now it is going to reiterate the 18:27contract. This is a contract. It writes, 18:29it's all writing all of this and it 18:31begins with opening statements. Begin 18:32out round one now. And it does. So the 18:34CEO sets the terms. The CTO says in line 18:38with his character, it's viable, but I 18:40worry. Enterprise pro prospects ask. You 18:42get the idea. It's now my move. I 18:44respond. I say, listen, we have to do 18:47this because we are losing in the 18:48market. A strong AI feature can add 18:50sizzle, and if we combine it with uh the 18:52CRO's concerns around SOCK 2, we should 18:54unstick some deals. It then moves 18:55forward. I'm not going to read all of 18:56these in detail, don't worry. We get to 18:58the second round, we see that it's 19:00playing the roles correctly. I then 19:02start to negotiate. I say, "Listen, we 19:05shouldn't just look at new customers. We 19:06should look at existing expansion 19:07revenue. I'm starting to move forward. 19:10This is third round. I've got to finish 19:11up." The CRO basically says, "Okay, if 19:13we're talking about expansion revenue, 19:15maybe I can get on board." The CFO who's 19:17been worried about the money says, 19:18"Okay, that that unlocks that for me, 19:21etc." And at the end, it gives me a 19:23scorecard. It says, "I got a partial 19:25approval." It gives itself a scorecard 19:27for twin realism and it gives me a 19:30debrief. What did it hinge on? Where did 19:32the tension surface? How did it work? I 19:34think this is the most useful part of 19:36this whole exercise because if you want 19:37to game out different scenarios like for 19:39a job interview, you get debriefs on 19:41your performance like this. I think it's 19:43really cool. Next, let's actually go 19:45into the job search arena and let's look 19:48at a real example of compensation 19:52negotiation, the final stage of job 19:54search. Okay, we are back. I'm not going 19:57to rerun this prompt in more detail. You 19:59get the idea. We went through the 20:00different situations. We're negotiating 20:02compensation this time. Smaller group, 20:04head of HR, CPO, and me. Uh, the win is 20:08dollars. We want more of them. Uh, we 20:10have six rounds. And I give it uh my 20:14current offer. Uh, I'm totally making up 20:16these numbers, but they're not 20:17completely implausible. 20:19And it says it's got it. I give it 20:22personality. The CPO is a hard-bitten 20:24guy. HR wants to follow policy. I 20:27realize this sounds right out of central 20:28casting, but we're having fun. Okay. And 20:30I give it instructions. Again, all of 20:31this is necessary to set up the fidelity 20:34of the digital twins world. And so it 20:35then starts to print out what it's going 20:37to do. It prints out all of the stuff 20:40it's working with, all the stuff it 20:42knows so far. Begins round one. It gives 20:44opening statements based on what I've 20:46said, what the CPO said, what HR has 20:48said, and then it asks me to go from 20:50here. Essentially, they say no. And so I 20:52say, "Hey, this reflects the market 20:54average. Like, this is legit." And so 20:56they start to shift a little bit. They 20:58bring up finance and approvals as an 21:00issue. Uh HR is really concerned about 21:03it. It's my move. So I say, "Okay, I 21:06want to play with equity and bonus a 21:08little bit. What if I drop the comp 21:09down?" This is actually how a lot of 21:12comp negotiations go. But we don't get 21:14the chance to play them out. The exact 21:16reason we have digital twins is so that 21:19we can do stuff like this in a 21:21controlled environment because I could 21:22simulate this whole thing again at a 21:24different number and walk out with a 21:26different sense of how the narrative 21:27went. Okay, so director of product, CPO 21:29and head of HR come back, they talk 21:31through that we're getting closer. I 21:33love that it's keeping a concession 21:35ledger so we can actually see how it's 21:37going. I wish that was true in real 21:38life. This is one way that it's easier 21:40in this uh in this simulated 21:42environment. Uh, we are arguing over the 21:45clause for equity. I can live with a 21:48clause for more equity that's dependent 21:50on period of employment, I say, but I 21:52don't think it's reasonable to make it 21:53dependent on product revenue goals if 21:55the CPO's equity is not similarly tied. 21:58And the CPO dodges the question as real 22:00CPOS tend to do and then moves on and 22:03says fine, we won't we won't make that a 22:04thing. We are now close to an agreement 22:06as you can see from here. 22:09And at the end of the day, the only 22:11thing that's an issue is what if finance 22:13pushes back? And I provide a savvy 22:16approach in terms of how finance pushes 22:18back. What do we do? How do we hand? And 22:20so then they come back and they say, 22:21"Okay, basically we've got a deal." I 22:23say, "Let's get a deal done." It then 22:25gives me a scorecard. It gives me comp. 22:27It gives me uh a debrief on how it went 22:31and it gives me like madeup next steps 22:34for what I would do. I think this is the 22:36most useful bit. It talks through how I 22:37handled finance resist. It talks through 22:39like the good the way I anchored on 22:41local market data and insisted on par 22:43with executive incentives was really 22:45smart. Um it's just going to give you 22:47feedback and you can you can ask it at 22:50that point will give me more critical 22:51feedback, right? Like push me harder and 22:53it will do that. Last thing I want to 22:55show you is what happens if we run this 22:59exact same scenario token for token. 23:01Same like same answers all the way 23:03through but we used a different model. 23:05So, we're going to go and use Chad 23:06GPT40. 23:08Okay, here we are. Chad GPT40 up here. I 23:11am not going to rerun. These are all 23:13exactly the same. Exactly the same. Set 23:15it up exactly the same. Yep. Great. Go. 23:19It then prints out 23:21the entire contract and begins. What I 23:24notice immediately with 40 is that 40 23:27has more personality. Look at the way 23:29the CPO talks. 23:30>> I've been in this game a long time, and 23:32I'll be straight with you. We think 23:33you're the right person, but we're not 23:35playing paying fantasy baseball 23:36salaries. That is 40 language. That is 23:39not 03 would never do that. So, the 23:41personality difference pops out right 23:43away. I come back as much as I can. I'm 23:45using token for token the same thing. 23:47So, we focus you on model differences. 23:50Um, the CPO continues to have 23:52personality. Head of HR continues to be 23:54very bland. And moving forward, I say, 23:58okay, I can, you know, negotiate. I can 24:00adjust. Again, I'm keeping this exactly 24:02the same as I did for the 03 example. 24:04The CPO then begins to praise me, which 24:06would not happen in real life and is one 24:08of our first indications that the model 24:10is not as smart as 03. This is a little 24:12bit out of character for a CPO, but he 24:15continues to be very, very chatty. I 24:17then give the same response. And what's 24:18interesting is I decided to give the 24:20same response to keep the model as clean 24:22as possible. But you'll notice that I 24:24gave that response when we were talking 24:26about clauses and 4 is not talking about 24:30clauses here. So this is me literally 24:33throwing a little bit of a curveball to 24:34keep the exact same token stream for the 24:38model to maintain the test as cleanly as 24:40possible. Okay, CPO comes back. Head of 24:42HR understands. And you know what's 24:45interesting? This is where the model 24:46really begins to diverge right here. 24:48They just agree. That did not happen 24:51with 03. One of my top takeaways looking 24:54at this example is that 4 shows its 24:57dumbness by being too agreeable in 25:00digital twin scenarios. If you were 25:02simulating this with 4, I worry you 25:05would walk away with a false idea of how 25:08tough these negotiations can be. I 25:10thought 03 did a much better job 25:12simulating how deep into the weed 25:14sometimes comp negotiations can go. And 25:17here I I just say I can get a deal done, 25:19right? like they've basically given me 25:20what I want. And if you walk out, you 25:22can see that I got more cash. The offer 25:24was at 190 and I walked out at 200 25:26instead of 194. Essentially, you can 25:28literally measure in dollars the 25:30dumbness of the model. 40 gave me $6,000 25:33that 03 was able to withhold for me 25:35because it was a sharper negotiator, 25:37which is actually better for you when 25:39you're looking to simulate. So, with 25:41that, there you go. We have three 25:44different actual conversation examples. 25:46We have a detailed breakdown of how to 25:47build a digital twin in chat GPT. You 25:51don't have to do anything. You don't 25:52have to run any code. You understand the 25:54principles behind that prompt. You 25:56understand how it hangs together. You 25:58understand how you could tweak it or 25:59adjust it for other scenarios. And the 26:01prompt itself is a super prompt, which 26:03means that you can run that prompt and 26:06give it other scenarios. Don't just do 26:08compensation. Don't just do product 26:10proposals. You can do it with sales 26:11negotiations. You can do it with job 26:13interviews. Anything that requires 26:16multiple parties to discuss and agree, 26:18you got options. You can run this prompt 26:21to simulate it. This is what I mean 26:22about the power of AI for digital twins. 26:25It is a huge deal. Most of us are 26:27sleeping on it. And I think part of it 26:29is it's been tricky to know how to 26:30prompt. And so this is my attempt to 26:32like bring the bridge over so that you 26:34can see like we're going to build the 26:35bridge. You can see how the prompt works 26:37and it's not a mystery anymore. Hope 26:39that was helpful. Tips.