WEBVTT 00:00:00.000 --> 00:00:03.180 Most people type eight words into ChatGPT. They 00:00:03.180 --> 00:00:05.320 copy the generic answer. Then they just close 00:00:05.320 --> 00:00:08.400 the tab. It saves a few minutes. But it doesn't 00:00:08.400 --> 00:00:11.759 change your income, Pete. But top operators, 00:00:11.759 --> 00:00:14.660 they treat AI entirely differently. They treat 00:00:14.660 --> 00:00:17.059 it like a business partner, one that works while 00:00:17.059 --> 00:00:20.600 they sleep. Welcome to the deep dive. Today we 00:00:20.600 --> 00:00:22.899 are exploring a really fascinating playbook. 00:00:23.070 --> 00:00:25.589 It was built from observing over 50 founders. 00:00:26.109 --> 00:00:29.149 We are breaking down custom AI business systems 00:00:29.149 --> 00:00:32.810 from multi -model arguing to automated finance 00:00:32.810 --> 00:00:35.170 reviews. Yeah, it completely reframes how we 00:00:35.170 --> 00:00:37.210 use these tools. I have to admit something upfront. 00:00:37.369 --> 00:00:40.409 I still wrestle with prompt drift myself. Getting 00:00:40.409 --> 00:00:43.570 the AI to actually stay on track is hard. OK, 00:00:43.710 --> 00:00:45.850 let's unpack this. It is a very common struggle. 00:00:46.009 --> 00:00:48.070 You want it to do the heavy lifting, but you 00:00:48.070 --> 00:00:50.329 have to build the system first. You have to fundamentally 00:00:50.329 --> 00:00:52.869 fix how you talk to the machine. You do this 00:00:52.869 --> 00:00:55.390 when you are awake. So it actually works when 00:00:55.390 --> 00:00:57.649 you're asleep. Exactly. Let's start with what 00:00:57.649 --> 00:01:00.210 we call the Google trap. We treat AI like a simple 00:01:00.210 --> 00:01:02.350 search engine. We expect absolute magic from 00:01:02.350 --> 00:01:05.129 incredibly short prompts. We hit Enter and get 00:01:05.129 --> 00:01:08.109 frustrated. Right. It is a massive misunderstanding. 00:01:08.689 --> 00:01:11.129 Think about bringing a brand new human advisor 00:01:11.129 --> 00:01:13.530 in. You bring them into your office. You look 00:01:13.530 --> 00:01:16.129 at them and just ask, Should I raise prices? 00:01:16.569 --> 00:01:18.989 Without giving them any context at all? Exactly. 00:01:19.290 --> 00:01:21.329 You give them zero background, they would completely 00:01:21.329 --> 00:01:23.730 fail or give a textbook answer. It sounds exactly 00:01:23.730 --> 00:01:26.909 like a generic Wikipedia summary. AI works the 00:01:26.909 --> 00:01:29.629 exact same way. The solution is providing the 00:01:29.629 --> 00:01:32.159 full context first. You have to give the model 00:01:32.159 --> 00:01:34.319 the messy background. You do this before you 00:01:34.319 --> 00:01:36.420 ask for any decisions. You share your current 00:01:36.420 --> 00:01:39.219 business stage. You give it last month's monthly 00:01:39.219 --> 00:01:41.420 recurring revenue. So you provide your current 00:01:41.420 --> 00:01:44.079 churn rate too. You list the options you're actually 00:01:44.079 --> 00:01:47.599 weighing. You share your 90 -day goals. And crucially, 00:01:47.739 --> 00:01:50.060 you show your real fears. Yes, that last part 00:01:50.060 --> 00:01:52.079 is so important. Let's pause on that for a second. 00:01:52.250 --> 00:01:55.250 Why does giving the AI our real fears actually 00:01:55.250 --> 00:01:57.670 change the output? Well, because it stops the 00:01:57.670 --> 00:02:00.230 AI from giving Wikipedia summaries. It forces 00:02:00.230 --> 00:02:02.769 the model to weigh emotional and business risk. 00:02:03.170 --> 00:02:05.930 It pulls from crisis management instead of generic 00:02:05.930 --> 00:02:08.550 theory. So provide the messy background, not 00:02:08.550 --> 00:02:10.979 just the final question. Right, once you do that, 00:02:11.240 --> 00:02:13.759 things definitely shift. But even with perfect 00:02:13.759 --> 00:02:17.599 context, there is a major risk. One single AI 00:02:17.599 --> 00:02:21.000 model still has dangerous blind spots. You really 00:02:21.000 --> 00:02:23.039 cannot trust the first answer. Confident models 00:02:23.039 --> 00:02:25.599 can be flat out wrong. They might miss something 00:02:25.599 --> 00:02:28.939 totally obvious about your market. This brings 00:02:28.939 --> 00:02:31.340 us to the three model sequence. You use three 00:02:31.340 --> 00:02:33.719 different models in sequence. Step one is the 00:02:33.719 --> 00:02:37.129 draft phase using Gemini. Gemini is best for 00:02:37.129 --> 00:02:39.689 fetching fresh web information. It gathers the 00:02:39.689 --> 00:02:42.349 raw materials you need quickly. Step two is the 00:02:42.349 --> 00:02:44.770 critique phase with Claude. Yeah, Claude plays 00:02:44.770 --> 00:02:47.229 devil's advocate here. It looks for logical flaws 00:02:47.229 --> 00:02:50.289 or generic text. It actively finds where the 00:02:50.289 --> 00:02:53.240 logic falls apart. Step three is the finalize 00:02:53.240 --> 00:02:55.879 phase with ChatGPT. You paste the original draft 00:02:55.879 --> 00:02:58.919 and quad's feedback in. You ask ChatGPT to rewrite 00:02:58.919 --> 00:03:01.180 it into a professional style. Let's look at a 00:03:01.180 --> 00:03:03.439 high stakes pricing strategy example. You use 00:03:03.439 --> 00:03:05.680 this sequence to battle test the logic. You do 00:03:05.680 --> 00:03:07.479 this before deciding on your praising tiers. 00:03:08.039 --> 00:03:09.620 But setting up three different models sounds 00:03:09.620 --> 00:03:12.430 exhausting. Aren't we just multiplying the hallucination 00:03:12.430 --> 00:03:15.169 risk by using three models? Uh, not if you do 00:03:15.169 --> 00:03:17.610 it correctly. They aren't collaborating to invent 00:03:17.610 --> 00:03:20.490 new facts. They are actively instructed to argue 00:03:20.490 --> 00:03:23.389 and fix each other's biases. It creates a robust 00:03:23.389 --> 00:03:25.710 system of checks and balances. Right. Pit the 00:03:25.710 --> 00:03:27.469 models against each other to catch their blind 00:03:27.469 --> 00:03:31.169 spots. Two -Sec Silence. Sponsor, replace holder. 00:03:31.319 --> 00:03:33.800 Getting models to argue is highly effective, 00:03:34.259 --> 00:03:36.680 but typing out all those rules every time is 00:03:36.680 --> 00:03:39.979 exhausting. Opening a blank chat over and over 00:03:39.979 --> 00:03:43.020 is slow. The system desperately needs memory. 00:03:43.319 --> 00:03:45.620 This is where workspaces come into play. We're 00:03:45.620 --> 00:03:48.360 talking about Claude projects and skills. Yeah, 00:03:48.419 --> 00:03:50.680 and we should probably define skills here. Reusable 00:03:50.680 --> 00:03:53.159 instruction packs the AI only loads when it needs 00:03:53.159 --> 00:03:56.129 them. Perfect. Early data shows a massive shift 00:03:56.129 --> 00:03:58.770 with these. Teams cut repetitive drafting cycles 00:03:58.770 --> 00:04:02.389 by up to 87%. They do this just by using these 00:04:02.389 --> 00:04:04.789 built -in skills. You completely stop re -explaining 00:04:04.789 --> 00:04:07.030 yourself to the machine. You need three crucial 00:04:07.030 --> 00:04:09.430 files to stop generic writing. The first one 00:04:09.430 --> 00:04:12.750 is the anti -AI style file. You have to ban phrases 00:04:12.750 --> 00:04:15.930 like, in today's fast -paced world, ban dive 00:04:15.930 --> 00:04:19.149 in and game changer, ban unlock the power of. 00:04:19.329 --> 00:04:21.810 Why do models love those phrases so much? They 00:04:21.810 --> 00:04:24.170 are statistically over -represented in the training 00:04:24.170 --> 00:04:26.990 data. The model just defaults to the most probable 00:04:26.990 --> 00:04:29.910 next word. The second file is the voice profile. 00:04:30.509 --> 00:04:33.209 This dictates short sentences, mostly under 14 00:04:33.209 --> 00:04:36.610 words. You use you way over I. And you absolutely 00:04:36.610 --> 00:04:39.230 ban hedging words like maybe or perhaps. Exactly. 00:04:39.829 --> 00:04:42.949 The third file is the fact dossier. This is the 00:04:42.949 --> 00:04:45.649 unchanging truth about your business, your exact 00:04:45.649 --> 00:04:48.569 prices, stats, and your audience. It's like stacking 00:04:48.569 --> 00:04:51.129 Lego blocks of data to build a custom brain. 00:04:51.389 --> 00:04:53.310 But let me ask you this. If I only have time 00:04:53.310 --> 00:04:55.689 to make one file today, which is the most critical... 00:04:55.689 --> 00:04:58.449 Without a doubt, the fact dossier, tone can always 00:04:58.449 --> 00:05:01.509 be tweaked later. But made -up numbers ruin your 00:05:01.509 --> 00:05:04.389 audience's trust instantly. Give the AI permanent 00:05:04.389 --> 00:05:07.569 memory so it actually sounds like you beat. Once 00:05:07.569 --> 00:05:10.110 the AI has your permanent memory locked in, you 00:05:10.110 --> 00:05:11.990 can take your hands off the keyboard entirely. 00:05:12.050 --> 00:05:13.850 You can do this for highly repetitive weekly 00:05:13.850 --> 00:05:17.430 tasks. Let's talk about AI business agents. An 00:05:17.430 --> 00:05:20.060 automated workflow that checks data... drafts 00:05:20.060 --> 00:05:22.579 text, and notifies you. Exactly right. You just 00:05:22.579 --> 00:05:25.300 reclaim 5 to 10 hours a week. Automating Monday 00:05:25.300 --> 00:05:28.300 loops or Friday inbox sweeps is powerful. You 00:05:28.300 --> 00:05:30.500 review the final output and approve it. This 00:05:30.500 --> 00:05:32.560 leads us seamlessly into something called vibe 00:05:32.560 --> 00:05:35.199 coding. Vibe coding changes how we build digital 00:05:35.199 --> 00:05:37.879 tools entirely. You describe what you want in 00:05:37.879 --> 00:05:40.779 plain English. You use tools like cursor, lovable, 00:05:41.079 --> 00:05:43.680 or bold. Let's talk about the Duolingo chess 00:05:43.680 --> 00:05:46.139 app example. It was built by just two people. 00:05:46.279 --> 00:05:48.500 They had no coding or chess experience at all. 00:05:48.639 --> 00:05:52.040 And they hit seven million daily active users 00:05:52.040 --> 00:05:55.019 in six months. The core lesson here is crucial. 00:05:55.500 --> 00:05:58.879 When AI fails, don't rate a fancier prompt. You 00:05:58.879 --> 00:06:01.540 feed it a database of real examples, like those 00:06:01.540 --> 00:06:04.959 real chess puzzles they used. Whoa! Imagine scaling 00:06:04.959 --> 00:06:07.079 to a billion queries without knowing how to code. 00:06:07.300 --> 00:06:09.399 What's fascinating here is the underlying lesson. 00:06:09.560 --> 00:06:12.180 The core point is mastering the iteration loop 00:06:12.180 --> 00:06:14.680 itself. Why shouldn't I wait until I fully understand 00:06:14.680 --> 00:06:17.000 a market before building? Because the new playbook 00:06:17.000 --> 00:06:19.519 is build while you learn. Shipping ugly and looping 00:06:19.519 --> 00:06:22.019 five times beats a six month polished launch. 00:06:22.660 --> 00:06:24.959 Every prototype teaches you faster than reading. 00:06:25.279 --> 00:06:28.339 Build ugly, test with real data, and automate 00:06:28.339 --> 00:06:32.319 the boring loops. Two secs silence. Building 00:06:32.319 --> 00:06:34.819 and automating is obviously incredibly powerful, 00:06:34.920 --> 00:06:37.750 but the system is not fully complete yet. You 00:06:37.750 --> 00:06:40.329 must rigorously review the final output. And 00:06:40.329 --> 00:06:42.310 you have to review the ultimate bottom line. 00:06:42.810 --> 00:06:44.990 Let's talk about using Gemini for content review. 00:06:45.529 --> 00:06:47.750 Don't just use AI to write more text. Use it 00:06:47.750 --> 00:06:50.730 to actually cut more. Ask Gemini to find the 00:06:50.730 --> 00:06:52.550 three weakest sections of a script. And write 00:06:52.550 --> 00:06:54.889 a better two -line hook instead. Right. Catch 00:06:54.889 --> 00:06:57.410 those structural problems early. Then we have 00:06:57.410 --> 00:06:59.410 the monthly finance check. Run this check on 00:06:59.410 --> 00:07:02.470 the 1st or the 15th. You drop revenue, costs, 00:07:02.750 --> 00:07:04.949 refunds, and churn into clod or perplexity. You 00:07:04.949 --> 00:07:07.189 ask what changed and which products improved 00:07:07.189 --> 00:07:10.370 margins. You ask what hidden risks exist right 00:07:10.370 --> 00:07:14.209 now. But remember, AI is an educator, not a licensed 00:07:14.209 --> 00:07:17.310 CPA. Don't ask it to pick stocks for you. Use 00:07:17.310 --> 00:07:19.529 it to turn a three -hour finance dread session 00:07:19.529 --> 00:07:22.189 into a 20 -minute reality check. Is it really 00:07:22.189 --> 00:07:24.829 safe to dump raw business financials into a chatbot? 00:07:25.110 --> 00:07:27.430 You use privacy -focused enterprise modes for 00:07:27.430 --> 00:07:29.959 this? And you scrub any personally identifiable 00:07:29.959 --> 00:07:33.399 info first. Focus on the ratios rather than exact 00:07:33.399 --> 00:07:35.579 bank account numbers. Exactly. Look at the margins, 00:07:35.639 --> 00:07:39.240 not the raw routing numbers. Use AI to map your 00:07:39.240 --> 00:07:43.019 money and slash fluff. But consult a pro. Eat. 00:07:43.579 --> 00:07:45.879 If we connect this to the bigger picture, the 00:07:45.879 --> 00:07:49.220 shift is profound. We are moving away from transactional 00:07:49.220 --> 00:07:52.360 AI usage. Right. One simple prompt giving one 00:07:52.360 --> 00:07:55.319 generic, boring answer. We are moving rapidly 00:07:55.319 --> 00:07:58.759 toward infrastructural AI. workspaces, multi 00:07:58.759 --> 00:08:01.300 -model reviews, permanent files, and automated 00:08:01.300 --> 00:08:03.699 agents. It really becomes the true backbone of 00:08:03.699 --> 00:08:05.639 the business. Remember the rule of compounding 00:08:05.639 --> 00:08:07.920 here. Do not try all nine steps this week. Pick 00:08:07.920 --> 00:08:11.120 just one single step to start today. If you write, 00:08:11.259 --> 00:08:13.319 build the three permanent voice files first. 00:08:13.759 --> 00:08:15.720 If you sell a product, set up with a monthly 00:08:15.720 --> 00:08:17.959 finance review. Do it consistently for 30 days 00:08:17.959 --> 00:08:19.800 before moving to step two. I want to leave you 00:08:19.800 --> 00:08:22.279 with a final reflection. If vibe coding and agents 00:08:22.279 --> 00:08:24.339 make generating content and code nearly free 00:08:24.339 --> 00:08:26.660 and instant for everyone, what happens to the 00:08:26.660 --> 00:08:29.279 value of human taste? When anyone can build an 00:08:29.279 --> 00:08:31.399 app in a weekend, the advantage shifts entirely 00:08:31.399 --> 00:08:33.559 from the speed of your typing to the quality 00:08:33.559 --> 00:08:35.720 of your curation. Something to think about. Out 00:08:35.720 --> 00:08:36.399 to your own music.