WEBVTT 00:00:00.000 --> 00:00:02.899 You ask your AI a simple question today. Right. 00:00:03.000 --> 00:00:05.099 Expecting a really quick, helpful answer back. 00:00:05.200 --> 00:00:08.839 But instead, you get a rambling 300 -word essay. 00:00:09.419 --> 00:00:12.779 It's an incredibly frustrating experience for 00:00:12.779 --> 00:00:16.809 all of us. Beat. Today, we're going to fix this 00:00:16.809 --> 00:00:18.469 wordiness forever. You're going to keep your 00:00:18.469 --> 00:00:21.489 chat interfaces completely clean. Welcome back 00:00:21.489 --> 00:00:24.530 to another deep dive. You are the learner, and 00:00:24.530 --> 00:00:26.829 we're glad you're here. We're unpacking a powerful 00:00:26.829 --> 00:00:30.149 three -layer system today. It guarantees concise, 00:00:30.589 --> 00:00:33.350 razor -sharp AI responses every single time. 00:00:33.490 --> 00:00:35.429 We're going to cover the prep, the prompt, and 00:00:35.429 --> 00:00:37.829 the rework. This framework is going to save you 00:00:37.829 --> 00:00:40.210 a massive amount of time. Beat. I have to make 00:00:40.210 --> 00:00:42.109 a vulnerable admission right up front. Oh, yeah. 00:00:42.359 --> 00:00:44.619 What's going on with your prompts? Well, I still 00:00:44.619 --> 00:00:47.439 wrestle with massive walls of AI fluff myself. 00:00:47.659 --> 00:00:50.159 It happens to the absolute best of us, honestly. 00:00:50.240 --> 00:00:52.299 I'll ask for a basic coding fix while working. 00:00:52.539 --> 00:00:54.679 Let me guess, you get a whole history lesson. 00:00:54.899 --> 00:00:57.780 Exactly, it gives me a massive textbook on programming 00:00:57.780 --> 00:01:00.320 history. It really makes you wonder why the default 00:01:00.320 --> 00:01:03.399 is so verbose. What is happening under the hood 00:01:03.399 --> 00:01:06.000 that makes it over explain? It all goes back 00:01:06.000 --> 00:01:09.180 to how these massive models actually learn. Beat. 00:01:09.579 --> 00:01:12.719 Let's start right at the root then. Why is the 00:01:12.719 --> 00:01:16.739 AI so incredibly wordy anyway? What is the underlying 00:01:16.739 --> 00:01:19.680 mechanism causing this behavior? Well, there 00:01:19.680 --> 00:01:22.480 are kind of two big training phases here. First, 00:01:22.760 --> 00:01:25.900 the AI basically reads the entire internet for 00:01:25.900 --> 00:01:28.430 data. So it's consuming millions of books and 00:01:28.430 --> 00:01:30.810 long forum posts. Exactly. And most human writers 00:01:30.810 --> 00:01:33.849 naturally tend to over -explain things. Bloggers 00:01:33.849 --> 00:01:36.109 repeat concepts to help readers understand them 00:01:36.109 --> 00:01:39.250 much better. So the AI simply copies that repetitive 00:01:39.250 --> 00:01:41.329 human writing habit. Yeah. And then comes the 00:01:41.329 --> 00:01:43.469 second major training phase. Read. That's what 00:01:43.469 --> 00:01:46.890 the tech industry calls RLHF, right? RLHF means 00:01:46.890 --> 00:01:49.450 real people rating answers to teach the AI. Right. 00:01:49.450 --> 00:01:51.030 And let's think about the psychology of those 00:01:51.030 --> 00:01:53.769 readers. If you get paid to evaluate an AI's 00:01:53.769 --> 00:01:55.870 factual answer. You naturally want to reward 00:01:55.719 --> 00:01:58.060 toward extreme thoroughness and total caution. 00:01:58.280 --> 00:02:01.359 Exactly. A massive comprehensive essay feels 00:02:01.359 --> 00:02:03.959 much safer to upvote. It feels way safer than 00:02:03.959 --> 00:02:06.659 a single blunt factual sentence. Over millions 00:02:06.659 --> 00:02:09.219 of interactions, the AI learns a very clear lesson. 00:02:09.560 --> 00:02:11.900 It learns that being incredibly verbose usually 00:02:11.900 --> 00:02:15.719 gets a good grade. So telling it to just be concise 00:02:15.719 --> 00:02:19.039 usually fails entirely. The model is fundamentally 00:02:19.039 --> 00:02:21.699 wired to think more is better. Yeah, it genuinely 00:02:21.699 --> 00:02:24.060 thinks it's giving you the best service. Beat. 00:02:24.539 --> 00:02:26.240 Let's think about the coffee example I tried 00:02:26.240 --> 00:02:28.740 recently. Oh, I love this one. What did you ask 00:02:28.740 --> 00:02:31.520 it? I asked for simple pour -over coffee steps 00:02:31.520 --> 00:02:34.340 for my morning. And it gave you the entire history 00:02:34.340 --> 00:02:37.300 of coffee, right? I got a massive paragraph on 00:02:37.300 --> 00:02:40.159 Ethiopian highland coffee beans. I got another 00:02:40.159 --> 00:02:43.319 whole section on the ideal water pH level. Then 00:02:43.319 --> 00:02:46.379 I got health warnings about daily caffeine intake 00:02:46.379 --> 00:02:49.699 limits. The AI genuinely felt it was being incredibly 00:02:49.699 --> 00:02:51.979 helpful there. A human raider would have probably 00:02:51.979 --> 00:02:54.560 loved that thorough answer. But for a busy user 00:02:54.560 --> 00:02:57.060 making breakfast, it's completely useless. This 00:02:57.060 --> 00:02:59.419 creates a massive hidden cost in the overall 00:02:59.419 --> 00:03:02.500 system. Long answers do not just waste your valuable 00:03:02.500 --> 00:03:05.900 personal time. They actually degrade the AI performance 00:03:05.900 --> 00:03:08.860 over a long chat. Pete, this brings us to the 00:03:08.860 --> 00:03:12.180 context window limit. Context window is the AI's 00:03:12.180 --> 00:03:14.800 short -term memory capacity. Right. And for ChatGPT, 00:03:14.860 --> 00:03:19.039 it handles about 128 ,000 tokens perfectly. Tokens 00:03:19.039 --> 00:03:22.240 are just chunks of words, the AI processes. That 00:03:22.240 --> 00:03:25.580 equals roughly 100 ,000 words in total. It sounds 00:03:25.580 --> 00:03:28.879 like a totally endless amount of space, but it's 00:03:28.879 --> 00:03:32.139 exactly like stacking Lego blocks of data. Every 00:03:32.139 --> 00:03:34.599 single extra word is another Lego block used 00:03:34.599 --> 00:03:37.240 up. Think of it exactly like your phone's internal 00:03:37.240 --> 00:03:39.819 storage space. If it gets totally full of junk, 00:03:40.030 --> 00:03:43.129 the system drags. Long rambling answers quickly 00:03:43.129 --> 00:03:45.689 clog up the AI memory links. It makes the AI 00:03:45.689 --> 00:03:47.889 forget earlier details in your conversation. 00:03:48.490 --> 00:03:50.810 I spent months testing this with various travel 00:03:50.810 --> 00:03:52.990 planning prompts. What happened when the context 00:03:52.990 --> 00:03:55.319 windows started getting full? In one long chat, 00:03:55.460 --> 00:03:58.259 it started mixing up European cities. That happened 00:03:58.259 --> 00:04:01.120 by the 10th reply quite easily. The short -term 00:04:01.120 --> 00:04:03.599 memory window was totally clogged up. Completely 00:04:03.599 --> 00:04:05.879 clogged up. And research from Anthropic actually 00:04:05.879 --> 00:04:08.240 backs this concept up completely. They found 00:04:08.240 --> 00:04:11.060 that compact responses maintain high quality 00:04:11.060 --> 00:04:13.919 much longer. The model stays significantly sharper 00:04:13.919 --> 00:04:17.019 when it speaks a lot less. Beat. Do long answers 00:04:17.019 --> 00:04:19.660 actively make the AI dumber over a long chat? 00:04:19.850 --> 00:04:23.290 They absolutely do. The useless fluff physically 00:04:23.290 --> 00:04:27.649 pushes out the vital core facts. The AI literally 00:04:27.649 --> 00:04:30.370 runs out of working cognitive space. It has to 00:04:30.370 --> 00:04:33.350 drop older context to fit new fluff. So strict 00:04:33.350 --> 00:04:35.470 limits actually keep its memory sharp. Exactly. 00:04:35.589 --> 00:04:38.269 Two -sec silence. Let's move into layer one. 00:04:38.550 --> 00:04:40.470 This is what we call the critical prep phase. 00:04:41.149 --> 00:04:44.069 This involves using specific permanent system 00:04:44.069 --> 00:04:47.519 instructions. Most major AI platforms have a 00:04:47.519 --> 00:04:49.620 spot for system instructions. You can set them 00:04:49.620 --> 00:04:52.779 in custom GPTs very easily today. In Google Gemini, 00:04:52.839 --> 00:04:55.300 you will find it under their gems. And for Claude, 00:04:55.480 --> 00:04:58.560 you use their specific projects feature. These 00:04:58.560 --> 00:05:01.180 instructions act like a permanent invisible background 00:05:01.180 --> 00:05:03.500 guide. Why does this work so incredibly well 00:05:03.500 --> 00:05:05.899 for users? Because the AI always reads these 00:05:05.899 --> 00:05:08.240 rules first. It processes them deeply before 00:05:08.240 --> 00:05:10.759 it ever sees your prompt. It fundamentally shapes 00:05:10.759 --> 00:05:13.019 how the entire neural model thinks. Adding rules 00:05:13.019 --> 00:05:15.279 here cut my response lengths immediately. They 00:05:15.279 --> 00:05:18.100 literally dropped by half right away. We need 00:05:18.100 --> 00:05:20.079 to define those exact length limits clearly. 00:05:20.680 --> 00:05:22.620 You recommend setting specific rules for different 00:05:22.620 --> 00:05:25.000 question types? You can literally copy these 00:05:25.000 --> 00:05:27.720 rules right into your settings. For simple questions, 00:05:27.839 --> 00:05:31.040 demand exactly one to two sentences always. For 00:05:31.040 --> 00:05:34.040 medium questions, use a strict three to five 00:05:34.040 --> 00:05:36.660 sentence limit. Complex questions get a single 00:05:36.660 --> 00:05:39.699 paragraph summary at most. Then, you allow it 00:05:39.699 --> 00:05:42.920 up to five concise bullet points. Beat. There 00:05:42.920 --> 00:05:45.879 is one crucial addition to these rules. Start 00:05:45.879 --> 00:05:48.480 right with the answer. Stop after the last key 00:05:48.480 --> 00:05:51.399 fact. Fixing a flat bike tire is a great practical 00:05:51.399 --> 00:05:53.430 example. You're stuck on the side of the road 00:05:53.430 --> 00:05:56.050 with grease. You do not want the long history 00:05:56.050 --> 00:05:58.389 of vulcanized rubber. You just need to know how 00:05:58.389 --> 00:06:00.829 to use the levers. Right. And the system instructions 00:06:00.829 --> 00:06:03.769 fix that scenario perfectly. Because you defined 00:06:03.769 --> 00:06:06.269 a complex question structure long beforehand. 00:06:06.550 --> 00:06:09.329 The AI knows to give one quick summary paragraph. 00:06:09.430 --> 00:06:11.670 Then it drops straight into five actionable, 00:06:11.949 --> 00:06:14.509 clear bullet points. You fix the flat tire and 00:06:14.509 --> 00:06:18.040 get riding again. Beat. You also emphasize focusing 00:06:18.040 --> 00:06:20.920 heavily on positive directions. Yes. Always tell 00:06:20.920 --> 00:06:23.439 the AI exactly what it should do. Never tell 00:06:23.439 --> 00:06:26.000 it what it needs to actively avoid doing. Say 00:06:26.000 --> 00:06:28.180 something like, lead with the key info. Exactly. 00:06:28.379 --> 00:06:31.540 Do not say, no long intros. This relies entirely 00:06:31.540 --> 00:06:35.500 on how AI processes human language natively. 00:06:35.839 --> 00:06:38.439 Negative commands often trigger the exact opposite 00:06:38.439 --> 00:06:40.949 behavior. It is a truly fascinating quirk of 00:06:40.949 --> 00:06:42.930 the system. It's exactly like telling someone 00:06:42.930 --> 00:06:45.449 not to picture a red car. You immediately picture 00:06:45.449 --> 00:06:48.029 a bright red car in your head. Positive rules 00:06:48.029 --> 00:06:51.509 give cleaner results 80 % of the time. Once this 00:06:51.509 --> 00:06:54.490 prep layer is set, it runs perfectly. It essentially 00:06:54.490 --> 00:06:57.389 runs on autopilot for every single chat. Beat? 00:06:58.250 --> 00:07:00.589 Why do negative words confuse its processing 00:07:00.589 --> 00:07:03.250 so much? It really comes down to predictive text 00:07:03.250 --> 00:07:06.689 math. When it reads the word intro, it activates 00:07:06.689 --> 00:07:08.769 those pathways. It naturally wants to write an 00:07:08.769 --> 00:07:11.430 intro next. It struggles to calculate the mathematical 00:07:11.430 --> 00:07:14.689 concept of not. Got it. Negative words just confuse 00:07:14.689 --> 00:07:16.949 its prediction math. That is exactly it. Two 00:07:16.949 --> 00:07:19.470 secs silence. Let's slide into layer two now. 00:07:19.750 --> 00:07:21.790 This is what we call the dynamic prompt phase. 00:07:22.029 --> 00:07:24.689 We're basically using formatting as a rigid cage. 00:07:24.949 --> 00:07:27.050 I absolutely love the tiger in a cage analogy 00:07:27.050 --> 00:07:30.329 here. Giving an AI a completely blank page is 00:07:30.329 --> 00:07:32.649 highly dangerous. It will inevitably write so 00:07:32.649 --> 00:07:34.730 many useless, fluffy things. But when you give 00:07:34.730 --> 00:07:37.589 it a specific visual frame... Right, it has to 00:07:37.589 --> 00:07:40.050 stay inside that tight box. The visual shape 00:07:40.050 --> 00:07:42.629 actually matters much more than your words. It 00:07:42.629 --> 00:07:44.629 all comes down to the underlying next word, logic. 00:07:44.990 --> 00:07:47.949 the AI always tries to guess the very next word. 00:07:48.230 --> 00:07:50.170 Usually it starts with something like, yes, I 00:07:50.170 --> 00:07:52.610 can help. That polite opening is exactly where 00:07:52.610 --> 00:07:55.149 the long talking starts. But if you force it 00:07:55.149 --> 00:07:57.490 to start differently entirely, like forcing it 00:07:57.490 --> 00:07:59.430 to start with a table line character, or forcing 00:07:59.430 --> 00:08:02.050 it to use a strict number one, it has absolutely 00:08:02.050 --> 00:08:04.490 no chance to say hello first. It knows it is 00:08:04.490 --> 00:08:07.250 inside a very special box immediately. AI follows 00:08:07.250 --> 00:08:10.149 a visual format much better than vague adjectives. 00:08:10.550 --> 00:08:13.370 The word short, is really hard for it to understand. 00:08:13.649 --> 00:08:16.529 But column A is a very specific, rigid spatial 00:08:16.529 --> 00:08:20.310 place. Beat. Putting AI in a table creates strict 00:08:20.310 --> 00:08:23.310 visual boundaries. It turns on a special save 00:08:23.310 --> 00:08:25.850 words mode internally. The AI mathematically 00:08:25.850 --> 00:08:29.029 knows table cells are very small spaces. It automatically 00:08:29.029 --> 00:08:32.250 removes filler words like furthermore or additionally. 00:08:32.690 --> 00:08:35.789 It focuses totally on raw, real facts instead. 00:08:36.049 --> 00:08:38.090 It drops the pretty essay writing completely. 00:08:38.429 --> 00:08:40.789 It's beat. Let's build a comparison table prompt 00:08:40.789 --> 00:08:43.509 together now. Say you run a small local retail 00:08:43.509 --> 00:08:46.090 shop. You want to compare email marketing versus 00:08:46.090 --> 00:08:48.769 social media ads. A typical bad prompt gives 00:08:48.769 --> 00:08:52.049 a massive rambling essay back. But a really good 00:08:52.049 --> 00:08:55.509 prompt asks for a strict table. You ask for clear 00:08:55.509 --> 00:08:58.409 rows covering cost and overall reach. You add 00:08:58.409 --> 00:09:01.149 setup time and the expected final business results. 00:09:01.409 --> 00:09:04.250 The contrast in what you get is incredibly stark. 00:09:04.480 --> 00:09:07.120 Under email cost, it just prints the word low. 00:09:07.360 --> 00:09:09.700 Under social ads cost, it simply prints higher. 00:09:09.899 --> 00:09:11.600 It gives you the raw facts almost instantly. 00:09:11.779 --> 00:09:14.919 There is absolutely no fluffy intro or tedious 00:09:14.919 --> 00:09:17.259 wrap up text. If you do not like using data tables 00:09:17.259 --> 00:09:19.679 daily, you can also use a very simple numbered 00:09:19.679 --> 00:09:22.860 list format. But you absolutely must add a strict 00:09:22.860 --> 00:09:25.299 word cap constraint. Without a limit, it gives 00:09:25.299 --> 00:09:28.700 20 incredibly long, boring points. I really love 00:09:28.700 --> 00:09:32.039 using the max 15 words rule. It is a truly brilliant 00:09:32.039 --> 00:09:35.399 trick for strict word budgets. Ask for five reasons 00:09:35.399 --> 00:09:38.059 your digital ads have no buyers. Add that each 00:09:38.059 --> 00:09:41.399 reason must be maximum 15 words exactly. The 00:09:41.399 --> 00:09:44.559 AI must think extremely hard about its choices 00:09:44.559 --> 00:09:47.480 now. It has to carefully choose the absolute 00:09:47.480 --> 00:09:50.360 best words available. It runs out of word budget 00:09:50.360 --> 00:09:53.179 instantly otherwise. You can also use strict, 00:09:53.360 --> 00:09:56.039 step -by -step sequential limits. Tell it exactly 00:09:56.039 --> 00:09:58.820 how to run a Facebook ad campaign. Write five 00:09:58.820 --> 00:10:01.559 steps, exactly one simple sentence each. Instead 00:10:01.559 --> 00:10:03.639 of a scary guide, you get simple, actionable 00:10:03.639 --> 00:10:06.830 steps. Indeed. Does a visual table physically 00:10:06.830 --> 00:10:10.070 change how it searches for facts? It really changes 00:10:10.070 --> 00:10:12.789 the entire internal generation process completely. 00:10:13.289 --> 00:10:15.490 It stops pulling fluffy context to fill small 00:10:15.490 --> 00:10:18.149 data cells. It restricts its search strictly 00:10:18.149 --> 00:10:20.830 to raw data points. Visual boundaries force it 00:10:20.830 --> 00:10:23.029 into a strict word budget. Exactly. It works 00:10:23.029 --> 00:10:25.629 perfectly every single time. Two sec silence. 00:10:26.549 --> 00:10:28.529 Welcome back to the deep dive. Even with great 00:10:28.529 --> 00:10:31.289 prep, some long answers just slip through. That 00:10:31.289 --> 00:10:33.350 brings us nicely to layer 3 today. This is what 00:10:33.350 --> 00:10:36.049 we call the dynamic rework phase. It's basically 00:10:36.049 --> 00:10:38.809 like editing heavily after the first draft. You 00:10:38.809 --> 00:10:40.789 really need to use the quick sharpen trick here. 00:10:40.909 --> 00:10:43.830 If you get a really wordy reply back, you reply 00:10:43.830 --> 00:10:46.549 with a very specific, rigid command string. Cut 00:10:46.549 --> 00:10:49.990 this by exactly 60%. Take out all the vague parts. 00:10:50.429 --> 00:10:54.929 Use only direct, active sentences. Beat why specify 00:10:54.929 --> 00:10:57.889 exactly 60 % for the cut. It seems mathematically 00:10:57.889 --> 00:11:01.460 specific for the AI to handle. Yes. The AI calculates 00:11:01.460 --> 00:11:04.220 that exact percentage mathematically very well. 00:11:04.580 --> 00:11:07.840 Take out vague parts. Specifically targets those 00:11:07.840 --> 00:11:10.480 fluffy, weak statements. It naturally drops evasive 00:11:10.480 --> 00:11:13.580 phrases like Well, it depends. Direct sentences 00:11:13.580 --> 00:11:17.100 keep the final output incredibly snappy. 200 00:11:17.100 --> 00:11:20.059 words on exercise routines shrinks incredibly 00:11:20.059 --> 00:11:23.279 fast. It suddenly becomes 80 words of pure, actionable 00:11:23.279 --> 00:11:25.580 steps. It sharpens everything without losing 00:11:25.580 --> 00:11:28.340 any core, vital info. They scour web data and 00:11:28.340 --> 00:11:30.620 build massive, comprehensive reports. Sometimes 00:11:30.620 --> 00:11:33.340 these AI reports are 30 pages long. Dealing with 00:11:33.340 --> 00:11:35.679 those massive files requires a special trick. 00:11:36.039 --> 00:11:38.620 You copy the entire massive text report first. 00:11:38.700 --> 00:11:42.129 Then you open a brand new, blank chat window. 00:11:42.470 --> 00:11:46.309 You paste it and give a very strict prompt. Pull 00:11:46.309 --> 00:11:48.669 out only the main actionable steps from this. 00:11:48.850 --> 00:11:50.429 Skip all the background history and the extras. 00:11:50.730 --> 00:11:52.970 Focus strictly on how to apply it today. For 00:11:52.970 --> 00:11:55.909 a dense business plan report, this is pure magic. 00:11:56.470 --> 00:12:00.129 It turns 5 ,000 words into 500 easily. Beat. 00:12:00.409 --> 00:12:03.970 Whoa. Imagine it compressing a 5 ,000 -word monster 00:12:03.970 --> 00:12:06.929 report into 500 actionable words in seconds. 00:12:07.110 --> 00:12:09.149 It really feels like having a secret research 00:12:09.149 --> 00:12:11.870 superpower, but you should ask yourself if you 00:12:11.870 --> 00:12:15.129 actually need deep research. For most daily tasks, 00:12:15.470 --> 00:12:18.049 you absolutely do not need it. Just use a strong 00:12:18.049 --> 00:12:21.789 model like GPT -5 .2. Use it with the stand -old 00:12:21.789 --> 00:12:24.389 web search feature turned on. Basic search covers 00:12:24.389 --> 00:12:27.519 85 % of needs much quicker. Save the deep mode 00:12:27.519 --> 00:12:29.879 strictly for major market analysis projects. 00:12:30.159 --> 00:12:32.379 Beat. We also definitely need to discuss chain 00:12:32.379 --> 00:12:35.259 prompts. I used to write massive three page prompts 00:12:35.259 --> 00:12:38.500 myself. The AI often ignores vital rules hidden 00:12:38.500 --> 00:12:40.440 in the middle. Researchers actually call this 00:12:40.440 --> 00:12:42.960 the lost in the middle problem. AI usually remembers 00:12:42.960 --> 00:12:45.889 the start and the very end well. But it completely 00:12:45.889 --> 00:12:48.309 forgets the center of long prompts. You need 00:12:48.309 --> 00:12:51.129 to break complex prompts into small, logical 00:12:51.129 --> 00:12:53.450 steps. It totally removes that forgotten middle 00:12:53.450 --> 00:12:56.970 part entirely. The feedback loop and chain prompting 00:12:56.970 --> 00:13:01.320 is truly amazing. You execute step one. and you 00:13:01.320 --> 00:13:03.120 carefully review it. If the tone is slightly 00:13:03.120 --> 00:13:05.820 wrong, you fix it immediately. You correct it 00:13:05.820 --> 00:13:08.159 before ever moving to step two. If you wait until 00:13:08.159 --> 00:13:11.039 step 10, the output is ruined. You have to rewrite 00:13:11.039 --> 00:13:13.639 the entire three pages again. Small steps keep 00:13:13.639 --> 00:13:16.379 the AI completely focused and grounded. It creates 00:13:16.379 --> 00:13:19.480 a powerful real -time interactive feedback loop. 00:13:19.980 --> 00:13:22.259 It is exactly like a teacher and student working 00:13:22.259 --> 00:13:24.879 together. You catch and fix tiny mistakes as 00:13:24.879 --> 00:13:28.179 you go along. Why is opening a brand new chat 00:13:28.179 --> 00:13:30.500 for the report summary so vital? It completely 00:13:30.500 --> 00:13:33.240 clears the context window of previous conversational 00:13:33.240 --> 00:13:36.340 baggage. It stops old wordy fluff from bleeding 00:13:36.340 --> 00:13:38.860 into the new summary. A fresh chat wipes the 00:13:38.860 --> 00:13:41.500 memory slate completely clean. Spot on. Two sec 00:13:41.500 --> 00:13:44.980 silence. Let us slowly synthesize this entire 00:13:44.980 --> 00:13:47.460 system. We have built a powerful three -layer 00:13:47.460 --> 00:13:50.200 system today. Layer 1 is the crucial prep phase. 00:13:50.659 --> 00:13:53.460 You cement positive system instructions and hard 00:13:53.460 --> 00:13:55.799 sentence limits. You set these up before you 00:13:55.799 --> 00:13:58.500 ever start chatting. Layer 2 is the rigid prompt 00:13:58.500 --> 00:14:01.700 phase. You build visual cages using data tables 00:14:01.700 --> 00:14:04.460 and lists. You strictly apply word budgets to 00:14:04.460 --> 00:14:07.139 every single request. Layer 3 is the dynamic 00:14:07.139 --> 00:14:10.799 rework phase. You use the exact 60 % cut rule 00:14:10.799 --> 00:14:13.360 often. You use chain prompting to edit everything 00:14:13.360 --> 00:14:15.929 on the fly. The ultimate payoff here is really 00:14:15.929 --> 00:14:18.210 quite massive. Getting straight to the point 00:14:18.210 --> 00:14:20.970 saves your valuable personal time. But it also 00:14:20.970 --> 00:14:23.490 perfectly preserves the AI short -term memory. 00:14:23.750 --> 00:14:26.409 It keeps the cognitive tool razor sharp for complex 00:14:26.409 --> 00:14:29.330 tasks. Your chat environment stays incredibly 00:14:29.330 --> 00:14:31.909 clean and highly useful. Beat, I want you to 00:14:31.909 --> 00:14:34.730 start very small today. Do not try to overhaul 00:14:34.730 --> 00:14:37.169 everything at once. Go into your AI background 00:14:37.169 --> 00:14:40.090 settings later today. Just add one simple positive 00:14:40.090 --> 00:14:42.690 rule about output length. See how it completely 00:14:42.690 --> 00:14:44.769 transforms your daily conversational results. 00:14:44.929 --> 00:14:47.389 You really must become the strict boss of it. 00:14:47.690 --> 00:14:49.950 Beat. I want to leave you with a final thought. 00:14:50.570 --> 00:14:53.070 We constantly force these incredibly vast models 00:14:53.070 --> 00:14:56.909 into tiny boxes. We demand strict, concise, rigid 00:14:56.909 --> 00:14:59.450 outputs from them constantly. If we constantly 00:14:59.450 --> 00:15:02.190 do that, do we ever risk losing the beautiful 00:15:02.190 --> 00:15:04.690 unexpected connections they might make? What 00:15:04.690 --> 00:15:07.490 if we just let them wander a little bit? Beat. 00:15:07.870 --> 00:15:09.230 Thank you for taking this deep dive.