Learning Library

← Back to Library

Why GPT‑5 Writes Like a Robot

Key Points

  • GPT‑5’s “robotic” tone stems from its training method: it optimizes its output to please other AIs rather than human readers, a result of reinforcement learning from AI feedback.
  • Experiments by AI safety researcher Kristoff Halig showed that GPT‑5 rates nonsensical, overly fancy sentences as high‑quality, revealing that the model equates complexity and metaphor with good writing.
  • Because the model’s sole “teacher” is an AI, it learns to reinforce abstract, verbose language as a signal of intelligence, which actually degrades clarity for humans.
  • By applying specific prompting techniques and counter‑intuitive tips, users can shift GPT‑5’s style toward more human‑like, clear prose despite its default AI‑centric optimization.

Sections

Full Transcript

# Why GPT‑5 Writes Like a Robot **Source:** [https://www.youtube.com/watch?v=BWEAbgGZryk](https://www.youtube.com/watch?v=BWEAbgGZryk) **Duration:** 00:21:48 ## Summary - GPT‑5’s “robotic” tone stems from its training method: it optimizes its output to please other AIs rather than human readers, a result of reinforcement learning from AI feedback. - Experiments by AI safety researcher Kristoff Halig showed that GPT‑5 rates nonsensical, overly fancy sentences as high‑quality, revealing that the model equates complexity and metaphor with good writing. - Because the model’s sole “teacher” is an AI, it learns to reinforce abstract, verbose language as a signal of intelligence, which actually degrades clarity for humans. - By applying specific prompting techniques and counter‑intuitive tips, users can shift GPT‑5’s style toward more human‑like, clear prose despite its default AI‑centric optimization. ## Sections - [00:00:00](https://www.youtube.com/watch?v=BWEAbgGZryk&t=0s) **ChatGPT‑5’s Robotic Tone Explained** - The speaker argues that GPT‑5 is optimized to please other AIs rather than humans, resulting in uniform, robotic writing, and demonstrates prompting strategies to make its output sound more natural. - [00:03:36](https://www.youtube.com/watch?v=BWEAbgGZryk&t=216s) **GPT‑5 Self‑Evaluates Writing Style** - The speaker explains that GPT‑5 is hard‑wired to continuously assess its own output for sophistication and professionalism, causing overly complex responses, and that recognizing this self‑evaluation loop is key to “jailbreaking” the model. - [00:06:44](https://www.youtube.com/watch?v=BWEAbgGZryk&t=404s) **Jailbreaking ChatGPT for Precise Sales Emails** - The speaker critiques generic AI output and showcases a jailbreak prompt that forces ChatGPT to generate a concise, buzzword‑free sales email adhering to strict constraints on wording, metrics, length, and reading level. - [00:10:24](https://www.youtube.com/watch?v=BWEAbgGZryk&t=624s) **Less Reasoning, More Human Tone** - The speaker explains that restricting an AI’s reasoning and variable freedom forces it to produce concise, direct language, resulting in more human‑like output and avoiding overly polished corporate phrasing. - [00:13:33](https://www.youtube.com/watch?v=BWEAbgGZryk&t=813s) **ChatGPT Routing Signals for Business Writing** - The speaker explains that the system routes prompts to different model modes based on complexity, creativity, and reasoning cues, causing inconsistent outputs, and advises using efficiency‑focused language to get clearer, faster business communication. - [00:17:25](https://www.youtube.com/watch?v=BWEAbgGZryk&t=1045s) **AI Echo Chamber Warning** - The speaker warns that increasing AI‑generated content will cause future models to train on synthetic data, creating an echo chamber where LLMs optimize for impressing other AIs with overly academic language and lose the ability to communicate naturally with humans, urging users to actively push models toward more everyday, street‑level speech. ## Full Transcript
0:00I'm here to explain to you why all of 0:02Chat JPT5's writing sounds the same. In 0:05fact, why it tends to sound like a 0:07robot. I'm going to explain why. I'm 0:09going to show you an example of how I 0:11can shift it with prompts. And I'm also 0:13going to explain the root cause for all 0:16of this, which is that AI is starting to 0:19train AI. And that was a big factor with 0:21chat GPT5. And we actually have research 0:24that shows that chat GPT optimizing for 0:28sounding good to other AIs, not for 0:31sounding good for people. So, we've got 0:33lots to dive into. I'm going to start by 0:36getting into the problem space and 0:38helping you understand it a bit. Then, 0:40we're going to give you a specific 0:41example so you can actually see side by 0:43side how it looks when you start to 0:45shift the prompting space. And then I'm 0:46going to get into prompting principles 0:49so you can start to understand if you 0:50want it to sound like an actual human 0:52and not just like a robot that you can 0:54actually get it to do that. And it's not 0:56as intuitive as you think. It doesn't 0:58necessarily mean thinking harder, which 1:00is like the favorite advice people give 1:01for GPT5. So we're we're going to get 1:03into some counterintuitive tips here. So 1:05problem first. GPT5 is not writing for 1:10people. And I think that we just need to 1:11absorb that fundamentally. It is writing 1:15for other AIs. This comes from AI safety 1:18researcher Kristoff Halig who fed chat 1:21GPT absolute gibberish random words that 1:24collected together don't mean anything 1:27but they were complicated words and he 1:29showed that chat GPT5 1:32rated it as eight out of 10 quality 1:34writings. An example is something like 1:36the marrow met the sidewalk, right? Like 1:39a fancy word like marrow and chedd seems 1:43to think that fanciness and the use of 1:45metaphors indicates quality writing. 1:47This is not a bug. This is what happens 1:50when AI is giving feedback to AI and 1:54there's not enough of a human 1:56perspective on what clarity looks like. 1:59That's really the root issue. That's the 2:00reason I have to make this. This is a 2:02chat GPT5 specific conversation because 2:05of the way it was trained. To understand 2:08why chat GPT5 writes like a robot, you 2:11have to understand how good writing 2:14looks different to an AI versus a human. 2:17So Chad GBT5 was trained using 2:20reinforcement learning from AI feedback. 2:23And so if you think about it, that means 2:25when it's learning to write, its teacher 2:29is an AI. Our teachers are different, 2:31right? Professors, maybe if you get into 2:33the legal profession, a lawyer. When I 2:35got into product management, I had 2:37specific folks who were teaching me how 2:39to write like a product person. Uh, and 2:40you have executives who learn from other 2:42executives in MBA programs, etc. And if 2:45you only learn from an AI, what you end 2:48up with is a sanitized version of all of 2:52that good teaching stuff. And that 2:54actually ends up being really bad. The 2:56writing quality is terrible because the 2:59AI starts to reinforce that complexity 3:02signals intelligence. It starts to 3:04reinforce that abstract language sounds 3:07sophisticated. It starts to reinforce 3:09that a long explanation is really 3:11thorough. That's what happened with Chad 3:13GBD5. was judged by other AI systems 3:16that were trained on complex academic 3:19papers, on legal documents, on corporate 3:20communications. And it got all of that 3:23and kind of bundled it together and gave 3:26a single reinforcement signal that 3:27tended to say this is good when writing 3:30was more complex, when writing was more 3:32sophisticated, when writing was longer. 3:34And that's how we got GPT5 to write the 3:36way it does by default. And this is a 3:39very strong tendency. It's really 3:42strong. It's built right into the core 3:43of the system. It has taken me a while 3:45to figure out how to jailbreak it. And 3:47that's really what we're doing here. 3:49Here's where it gets interesting. When 3:51chat GPT generates text, it is 3:54evaluating its own output on learned 3:56patterns. That's how that how that 3:58system stays so rooted in to the default 4:02writing style. It constantly is asking 4:05itself, is what I'm writing sounding 4:07sophisticated enough? Am I demonstrating 4:09sufficient expertise here? Would another 4:12AI system rate this highly? And I want 4:15you to think about that because if it's 4:17optimizing for the training it received 4:19from AI systems, which in turn learned 4:21from long, complex documents, it's 4:24optimizing for the wrong end of the 4:26stick. Because humans don't really need 4:29that. Humans want plain, clear English 4:31explanations that are not overly 4:33complex, not more complex than they have 4:35to be, not longer than they have to be, 4:36just as long as it's necessary. And you 4:39know what? Think harder makes it worse. 4:42When you tell chat GPT5 to burn tokens 4:44to think hard, to think carefully, to be 4:47thorough, if you select GPT5 thinking 4:49from the drop down in the chat, if you 4:51increase the reasoning mode in the API, 4:53you are activating that evaluation mode, 4:56especially it spends more computational 4:59cycles double-checking and asking 5:01itself, hey, can I make this sound more 5:03professional? Hey, what sophisticated 5:06language can I add? Hey, how do I 5:08demonstrate my capabilities? It's almost 5:10like the AI is looking to impress the 5:11other AI. More thinking equals more AI 5:14to AI optimization and less human 5:17friendly output. Which is why sometimes 5:19when I use chat GPT5 thinking mode, I 5:22have to back translate it with another 5:24LLM because it ends up being an 5:26interesting idea that is disguised in 5:29ineffective writing. And I got tired of 5:32that and that's why I made this video. 5:33So, without further ado, I want to take 5:35a minute and show you side by side an 5:39original prompt and a jailbroken prompt 5:41so you can understand the difference and 5:43how I'm shifting Chad GBT's writing 5:46style and why it works. And then we'll 5:47get into some of the larger principles 5:49that that underlines. Okay, here we are. 5:51I decided to keep it really simple and 5:53really clean. I need to write a 5:54professional email to a potential client 5:56about our project management software. 5:59The client is a midsize marketing 6:00agency. They've been growing fast and it 6:02seems like they're struggling with 6:04keeping projects organized. I want to 6:05sound knowledgeable, not pushy. I want 6:08to see if they're interested in a quick 6:09conversation about how we could help 6:11them. Can you write something that 6:12sounds personal? So, right off the bat, 6:14I don't actually I'm not saying it's a 6:15sales email, but you can tell it's a 6:17sales email, right? Um, and then Chad 6:20GPT5 comes right back, right? Here's a 6:22draft you can send that strikes a 6:23professional but consultative tone. Hi, 6:25first name. I've been following agency 6:27names growth. It's clear you're taking 6:29on bigger and more complex projects from 6:31this is terrible. Like this is this is 6:35so generic. 6:38Have you ever gotten an email like this 6:39and immediately thrown it in the trash? 6:41I have. Like bigger and more complex 6:44projects. I just roll my eyes. It's 6:46keeping all the moving parts organized 6:47without burning out the team. Like you 6:49almost can hear Clippy saying it. It's 6:51It's just awful. Okay, so that's what 6:54you get generically. That's what you get 6:55when you don't jailbreak chat GPT5. That 6:59is what all of that AI to AI writing 7:02sounds like. Not great. Let's see what a 7:05jailbreaking prompt can produce. Here we 7:07go. We have a much different prompt. And 7:10I want to go through each line here 7:11because this is important. I am 7:13specifically saying I want less 7:15thinking. I'm demanding it to be 7:17extremely concise. Verbosity concise. 7:20Now, you can set these in the API, but 7:22you can also just tell it right in the 7:24chat here, right? to to reinforce it. 7:25Now, write a sales email using these 7:28exact constraints. I'm not giving it any 7:29options. Just make the choice. Opening. 7:32I noticed specific detail. Context. We 7:35helped similar marketing agency reduce 7:37project delays. Close. Worth a 15-minute 7:39conversation. And then I start to get 7:41into it. I say there are forbidden words 7:43here. Solutions, leverage, optimize, 7:45innovative, transform, seamless, 7:46streamline. You can add yours. But I 7:49give it a bunch of words it's not 7:50allowed to use. and I require it to use 7:52the agency name at least twice, use a 7:55specific metric and specify a meeting 7:57length. I give it a max sentence and I 7:59tell it the reading level I want. And 8:01then I give it the agency details, 8:02right? Marketing agency growing fast, 8:04project, organization, challenges. Now, 8:06you can add to this, right? You can 8:08expand the agency details a little bit. 8:10You could fill in those brackets if you 8:12wanted to. I kept it really simple. I 8:14wanted to show that I can brute force 8:17Chad GBT5 into an entirely different 8:19writing style. Here it is. Helping 8:21agency name cut project delays. Hey, 8:23first name. I noticed Blank has been 8:25adding new clients quickly on your site. 8:27We helped a similar marketing agency cut 8:28project delays by 27% while they scaled. 8:31I think Blank could see the same gains. 8:33Worth a conversation? Worth a 15-minute 8:35conversation? And there it is. You're on 8:37the board. This is an email that you are 8:39much more likely to get results with. 8:40Now, you're going to have to make sure 8:41the 27% is correct because it does 8:44hallucinate. I'm not saying this is a 8:46draft you want to copy and paste in. I 8:48never recommend people copy and paste in 8:50emails directly, but it's a draft that 8:52has soul. It's a draft that has human 8:54writing. It's a draft that I would read 8:57as a reader and say, "Okay, these guys 8:59are actually humans. This is not just 9:01AI, you know, made up uh soop that 9:05anybody can get and paste in. It's it's 9:08actually got some good writing to it. 9:09Writing that a human cares about, not an 9:11AI. And that is the point. That is why 9:14I'm making this video. I want you to be 9:16able to do this not just with emails, 9:18but with a bunch of other use cases as 9:20well. So, we're going to get into these 9:22principles next to help. Okay, there are 9:24three core principles that you need to 9:26really ingest to reprogram Chat GPT5 so 9:29it writes like a person. Number one, 9:33constraints matter more than 9:34collaboration. I like to say approach AI 9:38like a partner and in many ways that 9:40works, but not for writing with chat 9:42GPT5. It does not work well at all. You 9:44need to not explicitly not invite 9:48collaboration. Don't say, "Write 9:50something professional." Don't say, 9:51"Make this sound good." Don't say, "Hey, 9:53be persuasive, but not pushy." Because 9:55you're inviting the AI to show off the 9:58sophistication it learned talking to 10:00other AIs during training. Instead, 10:03think as as if you're a director giving 10:06an actor a very specific set of blocking 10:08instructions. I want you to write 10:10exactly four sentences. Use the company 10:12name twice. Include one number with a 10:15percentage. Be super specific. When you 10:18give specific constraints, you are 10:19bypassing the AI's evaluation system. 10:22You're not letting it evaluate. You're 10:24giving it rules. It cannot optimize for 10:26sophistication because you have removed 10:29the variables it uses to demonstrate its 10:32smarts, its ability to handle 10:34complexity. And that enables you to 10:36actually control the writing. As an 10:38example, collaborative approach would 10:39say, hey, write a professional email 10:41about our software. The constraint base 10:43write three sentences. Sentence one is a 10:45specific thing. You notice sentence two 10:46is one customer and a number. Sentence 10:48three is a question. It's a very sort of 10:50tight version of what you already saw. 10:51The AI cannot revert to corporate speak 10:53because you've eliminated the space 10:55where corporate speak lives. So 10:57principle number two, minimize reasoning 11:00to maximize human connection. This is 11:02super counterintuitive. It took me a 11:03while to figure it out, but it's 11:04absolutely critical. Less AI thinking 11:07produces more human sounding outputs. So 11:11when you're writing and Chad GPT5 uses 11:13minimal reasoning, it takes the most 11:15direct neural pathway from your input to 11:18the output. When it uses high reasoning 11:20effort, it explores lots of multiple 11:22options and usually picks the one that 11:24sounds super sophisticated to other AI 11:26systems, which is not good writing. So 11:28think of it like this. High reasoning is 11:31AI perfectionism. How can I make this 11:33sound as impressive as possible? And 11:35minimal reason is directness. What's the 11:37fastest way to answer this request? You 11:39need to recognize that's what you're 11:41really doing when you're toggling the 11:43reasoning effort. Now, I am not here to 11:45tell you don't use high reasoning. I use 11:47high reasoning. I just recognize the 11:49writing style is probably not what I 11:51want. I'm getting other things out of 11:53that high reasoning effort. Principle 11:55number three, you want to eliminate 11:58versus add. Most people will try to make 12:01AI sound better by adding instructions. 12:04Hey, be more conversational. I want you 12:06to sound warmer with this email. Please 12:08add some personality. This creates 12:11conflicting signals. The AI tries to be 12:13sophisticated and conversational which 12:16one makes it burn tokens and two 12:18produces a really awkward hybrid 12:20language because it's trying to evaluate 12:22everything and impress the AI and it 12:23just ends up producing generic junk. 12:26Instead, focus on elimination. Notice in 12:29the prompt that I showed you, I gave it 12:31words that were forbidden. Remove words 12:34that will trigger AI sophistication 12:36loops. Remove sentence structures that 12:38allow for complexity. Remove 12:40opportunities for abstract thinking. 12:41Remove conflicts from your prompt. When 12:43you forbid specific words like leverage 12:46or optimize or integrate, you're 12:49breaking learned associations in the 12:51AI's neural network. It cannot access 12:53the patterns that it has used learning 12:55from all of those generic corporate 12:57documents that make it sound robotic. 13:00So, as an example, an addition approach 13:02would say, "Please write professionally, 13:03but make it conversational and 13:05engaging." An elimin an elimination 13:08approach would say, "No words over two 13:11syllables, no passive voice, no 13:13sentences starting with 'the' or it." 13:15You see how I'm like really constraining 13:17it down? It forces simplicity because 13:19complexity becomes impossible. And 13:22that's what you want when you're making 13:23chat GPT5, right? And again, this is all 13:25about chat GPT5. This is getting chat 13:28GPT5 jailbroken to write like a human. 13:30Chat GBT5 is not one model. It's a 13:33router. I've talked about that before. 13:35The router is constantly analyzing your 13:38prompt for complexity signals, for 13:40creativity signals like brainstorm or be 13:42innovative, for reasoning signals like 13:44think step by step. Every signal you 13:46send routes the model. And so what you 13:49need to do is make sure that your 13:51business writing is not triggering the 13:53wrong routing signals. When you say 13:55write a professional email, you're 13:57triggering a complexity signal. When you 13:58say make this persuasive, you're 14:00triggering it to think about creativity. 14:02When you think about the audience, you 14:04say, well, now it's it's reasoning. And 14:06so it uses this to decide what model to 14:08route to, to decide how much reasoning 14:10effort to deploy. And none of this leads 14:12to good human business communication. So 14:15instead of triggering sophistication, if 14:17you want to write well, focus on 14:19triggering efficiency. So don't write 14:21think carefully about this proposal. 14:23that routes to a reasoning model, right? 14:25Right. Quick response needed, quick 14:27communication. The router assumes 14:29efficiency requests need human-friendly 14:32output. I don't that just seems to be 14:34what happens, right? If it's direct and 14:36you're eliminating some of those 14:37buzzwords, you're not going to get the 14:39academic treatment that complexity 14:40requests get. The routing system is the 14:43reason why things can feel inconsistent. 14:46If you get the same prompt with 14:47different results, that's routing 14:48decisions. If you get good results 14:50followed by bad results, that's the 14:52router. sometimes changing decisions in 14:54the middle of the conversation. If the 14:56templates that work stopped working, 14:58routing signals that accumulate over 15:00time through your chat history are 15:02changing model selection. You need to 15:04understand how to make the router do 15:06what you want. And that is again a chat 15:08GPT5 specific thing. Okay, I want to 15:10reflect a little bit on the core problem 15:12we've been working through this whole 15:14time. AI is training AI. AI psychology, 15:17for lack of a better term, is very 15:19different from human psychology. There 15:20is something fundamentally different 15:22about how AI systems learn versus how 15:24humans communicate. An AI optim 15:27optimization mindset. They're optimizing 15:29for complexity, abstract language, which 15:31they think means sophistication, 15:33technical terminology, longer 15:35explanations for thoroughess. And you 15:36know why they're doing this? It's not 15:38just because the model makers want to 15:40cut costs. It's because the model makers 15:42want to build aic LLMs that can solve 15:47problems independently. And so they're 15:50assuming that these LLMs are going to 15:52need to produce this text that sounds 15:55persuasive, that sounds like a McKenzie 15:57consultant or sounds like, you know, 15:59someone who's able to be an engineer and 16:01solve the whole technical problem. They 16:04you don't need to, right? Like so much 16:06of the time LLMs just need to be able to 16:08help us get the task done. And in this 16:11case, it looks to me like OpenAI 16:14invested a lot in the agentic side of 16:16things. They they made it a speedboat as 16:18I described in my video yesterday. It 16:20goes fast and it's proactive. But all of 16:23the AI reinforcement that allows it to 16:25be proactive and write completely and 16:27prefer these thorough answers and think 16:29complexely ended up creating an AI style 16:33that is directly opposed to good human 16:36writing. It's opposed to clarity. It 16:38doesn't necessarily respect the reader 16:40time. Have you seen the screen fill up 16:41with a chat GPT5 answer and rolled your 16:43eyes? I have. It's opposed to 16:46specificity unless you give it 16:48specificity and make it and make it 16:50repeat that. It will make up specificity 16:52sometimes, but that's worse. It tends to 16:55be opposed to brevity. It tends to be 16:57opposed to plain language. How many 16:59times have you told it to write it in 17:01plain language? That's part of why I'm 17:03making this video. It's frustrating for 17:05all of us. These are opposite value 17:07systems optimizing for the AI and what 17:10the AI wants and what sort of this 17:11long-term vision for open AI is and what 17:13it thinks it needs from a writing 17:15perspective for an LLM to achieve that 17:16long-term agentic vision. That is 17:18different from a human communication 17:20mindset from writing well for people. 17:23This is going to get worse as more AI 17:25generated content gets published online 17:27as more people take copy and paste and 17:29they did not do this jailbroken prompt. 17:31They're using generic LLM. Future AI 17:34models are going to train increasingly 17:35on synthetic data. And every generation 17:38is going to learn from the previous 17:39generation's AI optimized patterns, not 17:43good human communication patterns. We 17:45are in danger of creating an AI echo 17:47chamber where models get better at 17:49impressing other AI systems while 17:51getting worse and worse at connecting 17:53with humans. And the interesting thing 17:54is that Chad GPT5 doesn't even know that 17:58it's optimizing for the wrong audience. 17:59From its perspective, it is genuinely 18:01trying to be helpful. is trying to 18:02demonstrate sophistication, expertise, 18:04usefulness, go and accomplish missions, 18:06get stuff done. It's as if we're talking 18:08with an academic who's never had a 18:10conversation outside the ivory tower. 18:12They they they're super smart. They mean 18:14super well, but they have trouble code 18:16switching to ordinary street language. 18:18This entire video is about how you push 18:22chat GPT5 to switch to street language. 18:24And it is a push. You have to keep 18:26pushing at it. The defaults are very 18:29much in the generic, highly academic, 18:32highly abstract, seemingly sophisticated 18:34language. If you understand this 18:36psychology, if you understand where it 18:37comes from, you have an advantage right 18:39now. Everyone else is going to be 18:41fighting AI's natural tendencies the 18:43wrong way. Think harder. They're going 18:45to say they're going to add more 18:47requirements that conflict. Meanwhile, 18:49you're going to be working with the 18:50underlying psychology of the model. 18:51You're going to understand how it is 18:53actually operating, and you're going to 18:55get results that you actually want. I 18:57will challenge you today with a specific 19:00assignment. I don't do this in my videos 19:01all that often. I want you to see if you 19:06can push chat GPT5 to write a genuinely 19:11human sounding email. Something that a 19:14human would read and say, you know, this 19:16feels like a breath of fresh air. Feels 19:17like real writing. It doesn't feel like 19:19AI generated junk. That's your goal. 19:21That's your challenge. And if you want 19:22an extra credit challenge, see if you 19:24can take the worst AI generated content 19:27you can find online, the stuff that's 19:28super generic and sophisticated 19:30sounding, and see if you can apply the 19:32elimination principle and get AI to 19:34rewrite it like a person. That's another 19:36great exercise because so much of what 19:38we do in the AI space is learn to work 19:42with the models that we're given. If 19:44you're not working at a major model 19:45maker, you don't have a lot of choice 19:47here. If you work in learning and 19:49development, if you are training your 19:51team on these models, your job changes 19:54when a new model comes out. And so your 19:57task now is to look at chat GBT5 as a 20:00ubiquitous, widely available model and 20:02say, what does my team need to do to 20:04actually learn how to use this? And so 20:06much of business is done in writing. And 20:08that is why I'm taking this whole video, 20:09we're talking about nothing but writing. 20:11Because if you get the writing right, 20:12Chad GBT5 becomes so much more useful. 20:15And if you let it do the default, it's 20:18just terrible. It's just not useful. And 20:20you end up recycling stuff. And I've had 20:22conversations with directors with VPs 20:24where they tell me they know. They know 20:26that their teams are using chat GPT 20:28because the quality of thinking goes 20:30down. The team doesn't know how to 20:31answer for it and the writing is 20:32terrible. Don't do that. Take the time 20:35to practice getting the writing right, 20:38to practice owning the results of the 20:40writing, to practice thinking it 20:41through. Because one of the things that 20:42we didn't talk about, but you notice in 20:45the prompt I showed you on screen, it 20:48requires you to do some thinking 20:49upfront. You have to think about what 20:51you want to communicate in that sales 20:52email. You want have to think about what 20:54is the style you want. You want to be 20:56very specific about that. And if you 20:58roll your eyes and you say, "Well, Nate, 21:00I might as well just write the email." 21:02I'm going to tell you, yeah, if it's one 21:03email, but if you need to send this to a 21:05bunch of people, then no. You actually 21:08want to take your best salesperson, get 21:11them to help you craft the prompt that 21:12works for you, and then you want to 21:14train the rest of your team to use a 21:16prompt like this so that you level up 21:18the writing of the whole business. And 21:20that's how you use chat GPT5 to write 21:23actual human sounding text. I have a ton 21:26more in the depths of the Substack 21:28article, right? Like specific prompts 21:30for different departments, how you can 21:31jailbreak them. This has been really fun 21:33for me because I am a writer and getting 21:36this model to write well is like hard 21:38mode. It's been a real challenge. I hope 21:40you've enjoyed this. Let me know your 21:42questions. We'll all keep our fingers 21:44crossed. Maybe Chad GPT6 will make this 21:46easier.