Learning Library

← Back to Library

Switch Models, Prompt Smarter

Key Points

  • The video’s first goal is to steer users away from defaulting to ChatGPT‑4 and instead adopt stronger reasoning models such as GPT‑3.5, Claude Opus 4, or Gemini 2.5 Pro, which deliver better performance and tool‑use transparency.
  • After selecting a superior model, the second goal is to simplify prompting by focusing on a handful of evidence‑based, memorable techniques rather than overwhelming users with dozens of tips.
  • The presenter distilled core prompting principles by reviewing guides from Anthropic, Google, OpenAI, and third‑party sources, identifying the most reliable strategies that actually improve results.
  • Effective prompting in 2025 emphasizes leveraging the model’s reasoning capabilities—asking it to think step‑by‑step, generate multiple candidate answers, and then evaluate or compare them.
  • By consistently using these concise, tested methods with a better model, users can achieve measurable gains without needing to study extensive prompting literature.

Full Transcript

# Switch Models, Prompt Smarter **Source:** [https://www.youtube.com/watch?v=hMKRBldkWEk](https://www.youtube.com/watch?v=hMKRBldkWEk) **Duration:** 00:15:10 ## Summary - The video’s first goal is to steer users away from defaulting to ChatGPT‑4 and instead adopt stronger reasoning models such as GPT‑3.5, Claude Opus 4, or Gemini 2.5 Pro, which deliver better performance and tool‑use transparency. - After selecting a superior model, the second goal is to simplify prompting by focusing on a handful of evidence‑based, memorable techniques rather than overwhelming users with dozens of tips. - The presenter distilled core prompting principles by reviewing guides from Anthropic, Google, OpenAI, and third‑party sources, identifying the most reliable strategies that actually improve results. - Effective prompting in 2025 emphasizes leveraging the model’s reasoning capabilities—asking it to think step‑by‑step, generate multiple candidate answers, and then evaluate or compare them. - By consistently using these concise, tested methods with a better model, users can achieve measurable gains without needing to study extensive prompting literature. ## Sections - [00:00:00](https://www.youtube.com/watch?v=hMKRBldkWEk&t=0s) **Upgrade Model, Simplify Prompting** - The speaker explains the importance of moving to a superior AI model and introduces a concise, proven set of prompting principles that are easy to remember and apply without sifting through extensive guides. - [00:03:39](https://www.youtube.com/watch?v=hMKRBldkWEk&t=219s) **Three Prompting Strategies** - The speaker outlines three validated techniques for improving LLM performance: request tool-generated code for math, generate multiple responses for self‑consistency, and have the model create a step‑by‑step plan before execution. - [00:06:47](https://www.youtube.com/watch?v=hMKRBldkWEk&t=407s) **Prompt Engineering: Guardrails & Positioning** - The speaker outlines three essential prompting strategies—constructing comprehensive guardrails and edge‑case handling, positioning critical instructions in the first and last 10% of the prompt, and prioritizing negative over positive examples—to keep probabilistic models reliably aligned with user intent. - [00:09:57](https://www.youtube.com/watch?v=hMKRBldkWEk&t=597s) **Prompt Optimization via Uncertainty Checks** - The speaker outlines techniques—uncertainty probing, capability discovery, and self‑improvement loops—to refine prompts, expose hidden ambiguities, and align model responses with realistic abilities. - [00:13:03](https://www.youtube.com/watch?v=hMKRBldkWEk&t=783s) **Beyond ChatGPT-4: Prompt Best Practices** - The speaker outlines structural prompting guardrails, self‑consistency, tool use, and planning techniques—effective for newer inference models—and urges users to stop relying on legacy ChatGPT‑4. ## Full Transcript
0:00We're going to do two things today. 0:01Number one, we are going to help you 0:03into a better model. I sound like a used 0:05car salesman, but we're going to get you 0:07into a better AI model, and I'm going to 0:08explain why it matters. And number two, 0:10we're going to talk about the state of 0:12prompting, and how prompting with that 0:14better model is something that you can 0:16learn to do without reading the 0:17thousands of pages of tips and guides. 0:20And frankly, to get ready for this 0:21video, I did a ton of research on those 0:23prompts and guides. So, you're not 0:24missing out. I looked at Anthropic, I 0:26looked at Google, I looked at thirdparty 0:28guides, I looked at OpenAI's guides. I 0:30wanted to see what are the overall 0:32principles of prompting now that we have 0:33reasoning models that we can start to 0:36pull out and name and really drill on 0:38clearly so that there's just a few 0:40things you can learn and take away with 0:42you that are easy to remember so your 0:44prompting actually gets better. That is 0:46my goal with this video. I don't want 0:48you to remember a 25 things that you 0:50need to think about when you prompt. I 0:52want you to have a clear model you can 0:54pick and I want you to have very 0:56memorable prompting tidbits that like 0:59actually make a measurable difference 1:01that have been tested to work and that 1:02there's not very many of. So let's get 1:04into it. Number one, finding a better 1:07model. Don't use chat GPT. That is 1:10almost always what people use when they 1:12say they use AI and that goes for CEOs. 1:14I have talked with CEOs who think there 1:16is no better model than 40 because four 1:18is bigger than three. And chat GPT is 1:20obviously the best. Neither of those two 1:22statements is as true as you might 1:24think. Three better models you can pick. 1:26All much better than 40. Chat GPT03. 1:30It's a reasoning model. I don't care 1:31what the number says. It's a little 1:33colder personalitywise, but it does a 1:35lot of work. It's my daily driver. 1:37Claude Opus 4. Fantastic model. Great 1:40writer. It thinks things through. It's a 1:42really, really good reader and it 1:43exposes its tool use transparently. 1:45Gemini 2.5 Pro, very strong model. you 1:49can go and use it over uh in Google's 1:51Vertex or in a lot of other places now 1:52that they're starting to expose it. It's 1:54a thinking model, a reasoning model. It 1:56also does tool use. It has a nice big 1:58context window and it works fast. These 2:00are all great choices. I'm not here to 2:02give you an extensive discussion about 2:0403 versus Opus 4 versus Gemini 2.5 Pro 2:07because frankly that is a a 5% of the 2:10population problem. The 95% of the 2:13population problem is getting people to 2:14stop using 40. If if we just decided to 2:18use 03, our collective perception of 2:20what AI could do would significantly 2:22improve. Okay, that being said, once you 2:25were using a better model, a reasoning 2:27model, a model that takes its time to 2:29think, which is what you get with 03, 2:31what do you do to prompt in a way that 2:34makes sense in 2025? I want to give you 2:37a few evidence-based techniques that 2:39actually work. Number one, ask the model 2:42to generate a few responses and then 2:45check the responses for consistency. And 2:47I'm not saying a few different responses 2:48to different questions. I'm actually 2:50saying ask the model to generate 2:52optionality and then to check responses 2:54for consistency. And so basically an 2:57example of that would be saying, give me 2:59five ways that you could define the 3:01answer to this question. Whatever the 3:03question might be, maybe tell me what an 3:05amphibian is, right, for biology. Give 3:07me five different definitions and check 3:08them for consistency. If you're solving 3:10a coding problem, give me five possible 3:12solutions and check them for 3:13consistency. Having the model check its 3:16work across multiple options is really 3:19cheap because producing those new 3:21solutions is relatively easy for these 3:23reasoning models. And having them check 3:25their work ensures that you are actually 3:27getting the benefit of that reflection 3:29or inference boost that you get from a 3:31reasoning model. The second technique I 3:33want to call out is program of thought. 3:35ask the model to solve this with math or 3:39code. And so instead of saying like if 3:41you have a math problem instead of 3:42saying please explain how you solve this 3:45just say write a function to solve this 3:47suggest that it call a tool. These 3:50inference machines, these the 03, Opus 3:534, Gemini 2.5 Pro, they can all write 3:56Python code. They can write code to 3:58solve these math problems. And that 3:59makes them much more accurate with 4:01numbers. Call for the use of the tool in 4:05the prompt. It's called program of 4:07thought, whatever you want to call it. 4:09Like you're basically programming it to 4:10call the tool by asking for it. It's not 4:12that hard. Number three, plan and solve. 4:15Ask it to create a stepby-step plan for 4:19a particular task first and then you get 4:22into execution. I do this with writing 4:24all the time or with anything that I 4:26need to do from a software perspective. 4:28Create a step-by-step plan too. I just 4:31ask it to lay out its thinking and then 4:33I critique that thinking. It's a huge 4:35step forward. It makes a big difference. 4:37Plan and solve. Plan and solve. So 4:39already there's three things that you 4:41can remember, right? Ask it to generate 4:42multiple responses, which is easy to do. 4:44and then critique them for consistency. 4:46That's a self-consistency tip. Ask it to 4:49use tools to solve math problems. That's 4:51a program of thought tip. Ask it to plan 4:53and solve. By the way, I'm not making 4:56these up. These are actually validated. 4:59These are the creme de la creme. These 5:01are the examples that are standing out 5:03as I survey the use of reasoning models 5:06across thousands and thousands of pages 5:08of prompt engineering best practices. 5:11I'm trying to give you the stuff that is 5:12actually easy to remember, easy to 5:15implement, and that you will see real 5:18benefit from right away. I'm not trying 5:20to give you fancy magic words. I'm 5:23trying to give you ways to engineer and 5:24work with these machines. So, with that 5:27in mind, I want to go below the level of 5:29the prompt. And I want to suggest that 5:31you also need to understand a couple of 5:33structural principles for whatever 5:35prompt you're writing in 2025. Number 5:38one, understand that you really need to 5:40give guardrails and edge cases. You 5:42can't just prompt for the happy path. If 5:44unable to X, then do Y. Please privately 5:48reflect on X, then summarizes Y, then do 5:51Z. You need to give it ways to handle 5:55the content you're giving it explicitly. 5:58Too many people stop at that here's the 6:01actual instructions happy path and don't 6:03do anything to provide structure around 6:06how the prompt is executed or edge cases 6:08or guardrails. I have so many prompts 6:11that I've put together and shown that 6:13illustrate this but the core concepts 6:16remain really consistent and it works 6:18regardless of whether you're in a cloud 6:21model or if you're in an open AI model 6:23or Google model. You want to be clear 6:25with your constraints and edge cases. 6:26You want to be clear with how you handle 6:28fallbacks. You want to be clear with the 6:29output structure. The actual 6:31instructions just need to be very clear 6:34and live nested inside a format that 6:37allows the model to know what your 6:39intent is if things don't go perfectly 6:42well or if it starts to stray far a 6:44field from the actual instructions. 6:46Essentially, 6:47the actual instructions are the motor 6:50and you have to build the ship around 6:52the motor through the guardrails and 6:54edge cases and all of that to get where 6:56you want to go. Otherwise, if you just 6:57stick a motor in the water, what's it 6:59going to do? It's going to sink no 7:00matter how well it runs. You can have a 7:02perfectly good set of actual 7:04instructions and it just doesn't go 7:06anywhere because it doesn't teach the 7:08model how to handle all of the things 7:09that can go wrong when a probabilistic 7:11model tries to understand your intent 7:13like that. So, I always say look for 7:16guardrails and edge cases. Claude system 7:18prompt is 90% guardrails and edge cases 7:20by the way. 90%. Just think about that. 7:23Okay, that's principle one. If you're 7:25prompting in 2025, principle two, 7:27context positioning. Attention is not 7:29uniform. Put your critical instructions 7:32in the first 10% of the prompt. You can 7:34put examples, you can put data in the 7:35middle, and then you reiterate with key 7:37constraints at the end. You want to 7:39assume that the model needs to anchor on 7:41this prompt and it takes the first part 7:43and the last part of that prompt very 7:45seriously. Principle number three, 7:46negative examples are more critical to 7:49include than positive examples. It is 7:52good to include a positive examples. It 7:54is even more important to include don't 7:56do X examples. Show failure modes 7:58explicitly and say avoid it. For 8:00example, if you're writing and you want 8:02it to sound like it's not silly, you 8:04might say, "Never use phrases like 8:06inconclusion or never use phrases like 8:09wrapping all of this up." Whatever you 8:11want to avoid and then it will just do 8:13that. It will know what bad looks like. 8:15Now, you're not going to be able to 8:16describe it exhaustively and so you're 8:18going to need to couch those bad 8:19examples inside a general statement that 8:21says this is the overall behavior to 8:23avoid and here's an example. But it is 8:25really important to include those 8:27negative examples so you don't run into 8:28issues. So, let's just re sort of review 8:30where we've been at. We talked first 8:32about evidence-based techniques that 8:34really work like self-consistency where 8:36you make it generate multiple responses 8:38and make it self-consistent across those 8:40responses which helps it to reduce 8:43random incorrect answers. We talked 8:45about program of thought where you ask 8:46it to use a tool like Python to solve a 8:49problem. We talked about planning first 8:50and then solving. So, those are specific 8:53tips. We then talked about structural 8:55principles. uh the idea of having 8:56guardrails and edge cases, the idea of 8:58having context positioning, first 10%, 9:00last 10% of the prompt being more 9:02important. We talked about negative 9:03examples beating positives just now. I 9:05want to give you one big insight at the 9:09end here that really unlocks where 9:12prompting is at. The key insight is 9:14this. Models know themselves better than 9:17we know them at this point. And so part 9:20of what you're doing is you're using a 9:23technique called metaprompting to help 9:25the models reveal what they know to you 9:27in a way that's usable. Again, we go 9:30back to the fact that we started this 9:31video moving from 40 to 03. If you are 9:34not using a better inference model, this 9:36will not go as well. So go back, use a 9:38better influence model, and the 9:40inference model will help you when 9:41you're doing metaparrompting. Let me 9:43give you a few examples with specific 9:45prompt phrases that you can use that are 9:47in the spirit of metaprompting or 9:49getting the model to help you prompt 9:50itself. A self-improvement loop. Here's 9:53my current prompt. Just write it out. 9:55How would you improve this prompt to get 9:57better results from you? It will come 9:59back with an answer. Very simple. You 10:01can write that different ways, but 10:03oftent times it's going to tell you 10:05things it knows about prompting that you 10:06didn't know for free. Isn't that great? 10:09Number two, uncertainty check. What 10:11parts of this request are unclear or 10:13ambiguous? What assumptions are you 10:15making? What additional information 10:17would help you execute this prompt with 10:19more accuracy? This pushes the model to 10:21voice hidden uncertainties that it would 10:23otherwise infer. It helps to prevent 10:25overconfident hallucination. Capability 10:28discovery is another one. How would you 10:30approach this if you had no constraints? 10:32What would be your ideal process? What 10:34tools or information would help you? 10:36This reveals what the model thinks it's 10:39capable of, which by the way is not 10:40always true. But often times you get 10:42suggestions for approaches that you 10:44hadn't considered before. And the model 10:46will reveal to you an overall approach 10:48and desire for information that you can 10:50help it with. And then you can wind it 10:52back from imagined capabilities that 10:54aren't true and get to a point where 10:56it's actually a useful prompt that has 10:57all the information that it needs. See, 10:59even those things, a self-improvement 11:01loop, an uncertainty probe, just 11:02checking to see if things are unclear, 11:04checking for the edge of capabilities 11:06that models have. These are techniques 11:08to help the model help you. It helps the 11:10model understand what's what it's able 11:12to do in order to support you as you 11:14prompt. But they're not the only ones. 11:16There's other things that sort of fit in 11:18this meta prompting bucket. Explain your 11:20reasoning step by step. What parts are 11:22you most or least confident about? 11:24Sometimes they hide the reasoning. And 11:26sometimes it's actually good to have 11:28prompts that hide reasoning that that 11:29can make sense and not distra distract 11:31from the output. But when you want to 11:33diagnose it, even if the reasoning isn't 11:35pure in the sense that it's exactly what 11:37it thought, asking the model to go back 11:39and think step by step is still really 11:42helpful and it helps you understand what 11:44the model thinks about the work that 11:45it's done so far. You can also use 11:47really old human techniques and the 11:49model has read so much human literature 11:51it actually works. The Socratic method 11:53works. What why did you choose that 11:55approach? I will ask the model that. 11:57What alternatives did you consider? I 11:58asked the model that all the time. What 12:00would you change if I changed this 12:02constraint? I also ask that. It uncovers 12:05implicit assumptions from the model, 12:07which is really, really interesting. So, 12:09these are not random tips. I don't want 12:12you to walk away from this video and 12:13think it's a bucket of random tips and 12:16an ask to not use 40. Instead, what I am 12:20surfacing are the best practices that 12:23have popped up again and again and again 12:26and again in this survey of literature 12:28that I've done. And so, when I look 12:29across all of prompting literature, I 12:31see some of these themes emerge and I 12:34don't see them talked about as clearly 12:35as they need to be. And so, my goal with 12:37this video has to been to be crystal 12:40clear. You want to use meta prompting. 12:43You want to ask the model to help you 12:44prompt. There are documented ways to do 12:47that. the self-improvement loop, the 12:48uncertainty probe, checking for 12:51capabilities, making sure that you 12:53understand how to ask for an explanation 12:55of work, understanding how to ask for uh 12:59clear reasoning through the Socratic 13:00method. It's not just me saying that. 13:02I'm trying to actually lad up the best 13:03practices so they're accessible. The 13:05structural principles I called out 13:07around guardrails and edge cases, around 13:09context positioning, around adding 13:10negative examples, these are structural, 13:12too. Like, they're really helpful to 13:14have regardless of model. And the 13:16evidence-based techniques that work, 13:17they're evidence-based for a reason. 13:19They're repeated for a reason. You want 13:21to have clear self-consistency so that 13:24you can give the model the chance to 13:25generate lots of options and then 13:27reinforce the elimination of 13:29hallucinations by forcing it to make 13:31those options consistent. You want to 13:33give the model a chance to call tools, 13:35program of thought. You want to give the 13:37model a chance to plan first. And again, 13:39I'll go back to the very beginning. 13:41These work because we have inference 13:42models. And for so many people, the 13:45reason why all of these tips are 13:46frustrating and why people keep earning 13:48clicks making all of these tips in 13:50prompting technique articles on the 13:52internet is because everyone's using 13:54chat GPT40 and we're not talking about 13:56that as a problem. It is a problem if 13:59you are using a model that is last 14:02generation. It is hard to make the 14:04techniques I'm describing here that will 14:07inform the rest of 2025. Like this will 14:09be helpful for chat GPT5 too. It's hard 14:12to make those work if all you're doing 14:14is sticking with 40. 14:16So, please move beyond chat GPT40. Now, 14:20eventually chat GPT5 may come out and 14:22they will literally drop it from the 14:24menu and you can't use it anymore. And 14:26when that happens, that's fantastic 14:28because people are going to get access 14:29to a reasoning model they didn't know 14:30they had. But again, at that point, 14:32they're still going to need prompting 14:34techniques that work. They're still 14:35going to need to understand how to use 14:37these models usefully. 14:39I hope that this has been a helpful 14:42summation of the literature. Are there 14:44thousands of pages and many many many 14:46more things to discover? There are and 14:47I've written up more on this so that you 14:49can go more in depth. But from an 14:51overall survey perspective talking with 14:54hundreds of people looking across the 14:56population, this is what I would want to 14:58say to most people about AI. Basically, 15:01just a few techniques that will help you 15:03dramatically improve your prompting plus 15:05pick a better model. It's sometimes 15:07getting better at AI.