How to Prompt GPT-5

0:00Today on the AI Daily Brief, 11 GBPT5 prompting techniques to actually get the most out of this new model. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:16All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, Blitzy, Vanta, and Superintelligent. To get an ad-free version of the show, go to patreon.com slash ai daily brief. And if you are interested in sponsoring the show, shoot us a note at sponsors at ai daily brief dot AI. Today we are taking a distinct step into the practical when it comes to GPT-5, and perhaps unsurprisingly, it ended up going quite long, so today we will just have this main episode. We will be back with our normal headlines main episode division tomorrow, but for now, let's dig into these prompting techniques.

0:49We are now almost a week into the GPT-5 era, and it continues to be a hugely divisive model, which is not to say that a lot of the initial complaints haven't been addressed. As you heard earlier in the week, the 4.0 Rebellion was successful and that model is now available for people. We also have a new approach to the UX selector, which is giving people who want it more fine-grained control over which model they're using. And a number of the complaints around limitations for plus users have also been addressed, at least in the short term. None of which is to say, though, who's right when it comes to how good this model is?

1:20People are still struggling to figure out how to get it to do what they want it to do. And while some have just retreated back to the models that they have already built workflows for, others have spent a bunch of time trying to really understand how to get the most out of this particular model. Even OpenAI themselves released a prompting guide for GPT-5. And what's interesting about this is that even though we have been trending more and more away from having to be good at prompting, I've discussed, for example, on this show how many different tools are effectively taking your prompt and turning it into a better prompt, which is, for example, what Ideogram, who I used to make this image, does.

1:58And yet, GPT-5, at least initially, suggests that we are in for a resurgence of prompt engineering. The TLDR background is that GPT-5 seems to require some very specific types of prompting. And so what we're going to do today, rather than endlessly debate how good or how bad the model is, or discuss what its social ramifications are, I'm going to share 11 techniques that come from all sorts of different sources of people who have actually figured out how to get more value from GPT-5. This should be extremely practical and something that you can put into place right away. We're going to break it up into two main sections, foundations, and then a short section for agentic toggles.

2:38Now, as we dig into the foundations, the biggest themes to note is simply the fact that this model requires prompting, but the upside of that is that it requires prompting because it's so good at adherence. Shai Almo from OpenAI writes, GPT-5 is very steerable and extremely good at instruction following. That means you'll need to update your prompts to get the best out of the model and help it generalize broadly. Alex Duffy from Every, who's had a few more labs experimenting with this, says, GPT-5 is more steerable than any other frontier model. Prompts make or break your results. And indeed, when I asked folks to share their early experience with this, this exact thing came up a lot.

3:15Seth Cronin, for example, wrote, My take is that you have to be a lot more explicit with GPT-5 if you want a deeper output. O3 seemed to read between the lines and give you a bunch of useful stuff to work with as if it wanted to do more work. Which brings us to prompting technique number one. And again, if you feel like this is a little bit of a regression, you are not alone. But a lot of people's best recommendations so far after a few days with this model is to basically tell it to think and work harder. Kajetan Mastella writes, The GPT-5 authors said it themselves. append think harder or think deeper to the end of every prompt.

3:47Works like a charm. Basically telling it to think deeper is an alternative way to trigger higher reasoning. Now it should be noted now that they have updated the GPT-5 selector to include auto, which decides for you how long to think, fast, which gives instant answers, thinking mini, which thinks quickly, i.e. a little bit of thinking, thinking, which thinks for a long time for better answers, and then pro, which presumably is the most thinky of all the things. So maybe you could now assume that the UI is taking care of this. However, others are finding prompts to really push GPT-5 to work harder to actually work.

4:21Alex Duffy, again from Every, shared his UltraThink prompt. It reads, First, think deeply for five minutes. UltraThink at a minimum. If after five minutes you still don't have the optimal response, keep thinking until you do. About the best way to do this. HyperWrite CEO Matt Schumer had an even longer version of this. He tweeted, Here's my insanely powerful GPT-5 prompt. It guides GPT-5 to think for way longer than it normally would, putting up to 3x more time and effort into solving your task. And basically what he proposes is before you actually prompt your task, you add this block of text which says, and obviously I'm only reading a part of this, Ultra deep thinking mode, greater rigor, attention to detail, and multi-angle verification.

5:02Start by outlining the task and breaking down the problem into subtasks. For each subtask, explore multiple perspectives, even those that seem initially irrelevant or improbable. Purposefully attempt to disprove or challenge your own assumptions at every step. Triple verify everything. Critically review each step. Scrutinize your logic assumptions and conclusions, explicitly calling out uncertainties and alternative viewpoints. And so on and so forth. So again, foundational prompting technique number one, just tell it to think or work harder. Next up, we have the concept of planning phases. And this almost reminds me of the old think step-by-step kind of prompt that we use to prompt chain of thought before we had the reasoning models.

5:40And this one comes from one of the better prompting guides we've seen so far from Pietro Chirano, the CEO at Magic Path. He tweeted a few days ago, GBT-5 requires a different way of prompting. It's much more susceptible to instruction, especially style and tone, and it does better when provided with reasoning validation and planning sections. Now Pietro had access to GBT-5 for a few weeks before it was announced, and so has a few more cycles to have come to these conclusions. He actually talks about these explicit planning phases in two different ways. In his main section, he writes GPT-5 excels when given explicit planning phases.

6:14An example prompt he shares is, before responding, please 1. Decompose the request into core components. 2. Identify any ambiguities that need clarification. 3. Create a structured approach to address each component. And 4. Validate your understanding before proceeding. It seems as though so far, giving GPT-5 these sort of explicit planning phases makes sure that the model doesn't skip steps and actually does a more comprehensive job. Interesting, in his advanced technique section, when talking about tool implementation, he suggests giving the model a mental to-do list structure that could also work for this sort of planning.

6:46In that example, he asks the model to track progress with primary objective, subtask one, subtask two, validation step, and final review. So a similar principle here. Our next prompting technique is to be extremely explicit about what you want, and this comes up in every single person's discussion. We heard it over here from Seth on my post. It's the subtext of all these arguments that the model is extremely steerable. It's here once again in Pietro's guide, where his first core principle is GPT's susceptibility to instruction style. He writes, GPT-5 is highly responsive to how you structure your prompts.

7:23Be explicit about your tone and style. The model adapts strongly to the communication style you establish. Use consistent formatting. maintain uniform structure throughout your prompts, and define expectations clearly. GPT-5 performs better with well-defined parameters. This, by the way, was validated by OpenAI when they shared their prompting guide. They write, As our most steerable model yet, GPT-5 is extraordinarily receptive to prompt instructions surrounding verbosity, tone, and tool-calling behavior. GPT-5 follows prompt instructions with surgical precision, which enables its flexibility to drop into all types of workflows.

7:58However they continue, Its careful instruction following behavior means that poorly constructed prompts can be more damaging to GPT-5 than to other models. So explicitness really matters, but so does structure, which is our next prompting technique, a really foundational prompting reminder. If there is just one thing to take away from all of this, it's to be more conscientious of how you structure your prompts. Now I will admit that this is something that I have gotten fairly lazy around. So many of these models at this point, including, frankly, the last generation of OpenAI models, had gotten so good at kind of interpreting what you wanted or just getting you started in a way that you could iterate with that I had gotten pretty lazy around prompt structure.

8:42Which is not to say that everyone had. You might have seen over the last month or so a rise in people talking about something called JSON prompting. Go search JSON prompting on X and you will see a million posts like this one. Tom Croshaw01 writes, there's a prompting technique that makes AI respond exactly how you want every single time. It's called JSON prompting, and it's why enterprise AI apps never break. Regular prompts give you inconsistent outputs because AI doesn't know what structure you want. JSON prompts force the AI to follow a specific format every time. Structure equals predictability.

9:17Now he gives an example of a JSON prompt for something as simple as write me a viral X post. In it, he breaks down that viral X post into hook, attention-grabbing opening line, content type, a thread, a single tweet, or a story, the main message, the core value proposition, social proof, like numbers, credentials, and results, the pain point, the problem being solved, et cetera, et cetera. The point is you break this thing up into its constituent elements, and then you use JSON formatting, which is basically a particular type of structuring information in a way that computers theoretically have an easier time reading to give it all those details.

9:53Now, like I said, this has become a major trend lately. But the secret here is that it's really probably not about the JSON formatting itself. AI in Design writes, I'm 99.9999 % sure that a prompt being in JSON format does nothing compared to one that's not. The quote unquote secret here is most likely simply the fact that doing it in this way forces structure you can add in any other way just as well. If all the fields were just headings like this without the JSON structure, you'd get the same result. Using an arbitrary chosen JSON structure being better than regular text defies all logic. And if you need more evidence, Nolan McCallum, who does Applied AI at OpenAI, writes, guys, I hate to rain on this parade, but JSON prompting isn't better.

10:36It physically pains me that these types of posts are getting so much traction. I've actually done experiments on this, and Markdown or XML is better. JSON isn't token efficient and creates tons of noise and attention load with white space, escaping, and keeping track of closing characters. JSON also puts the model in an I'm reading and outputting code part of the distribution, which is not always what you want. However, he says, I agree that being specific and using structure in your prompts is a good thing. And so this is the point that I'm trying to come around to. The reason that people are having such success with JSON prompting is not the JSON formatting itself.

11:09It's the fact that that format is forcing them to add an immense amount of fine-grained detail, which means that the fidelity of the model's adherence is going way up because they're giving it more to work with. And that broad idea seems likely to be translating extremely well to GPT-5, where it appears that the more clearly you structure your prompt, the better results you're going to get. This is in fact an entire category of Pietro's prompting guide. He gives a number of different structured prompting techniques include the spec format, which is sort of like a non-JSON requiring JSON prompt if you've seen these posts on X.

11:42So for example, he suggests a definition of what exactly you want accomplished, when it's required, i.e. the conditions that trigger the behavior, a format and style for the output, a sequence or a step-by-step order of operations, what to avoid and what's prohibited, and how to deal with unclear inputs. He also suggests reasoning and validation steps that you should include in complex prompts. He suggests, for example, pre-execution reasoning, a planning phase, validation checkpoints, and a post-action review. The easy takeaway, and honestly, if you just get one thing from this episode, is just to tighten the structure of your prompts and think more conscientiously about your prompts and you're going to immediately likely get better results than if you're being more loosey-goosey.

12:23And like I said, this is a bit of a departure. Some of the other models were getting so good at interpreting that it's kind of weird to have to go back to this, but hey, if it gets better results and if ultimately GBT-5 is like a more powerful tool but that has a higher bar to entry, it's certainly plausible that it's going to be worth the trade-offs. Now, our next prompting technique, which once again has a few people that I've seen saying versions of this, is to ask the model to share its thought process. So, for example, Pietro gave an example of this in those reasoning and validation steps where he asked the model to explain its understanding of the task and its approach, but this is also in the OpenAI prompting guide as well.

13:00They write, prompting model to give a brief explanation summarizing its thought process at the start of the final answer, for example via a bullet point list, improves performance on tasks requiring higher intelligence. So basically it seems like you get better results if the model knows that it's going to have to share not just its answer, but how it got to that answer and why it chose the path that it did. And whereas the structure-it-all prompting tip is very broad and vague, this seems like a very easy-to-execute, simple sentence-to-add in every complex prompt that, at least according to OpenAI, is increasing GPT-5's performance.

13:58then generates and precompiles code for each task. Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Blitzy is providing a limited time, 30-day free proof of concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org.

14:29Visit blitzy.com and press book demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native. That's blitzy.com. As a founder, you're moving fast towards product market fit, your next round, or your first big enterprise deal. But with AI accelerating how quickly startups build and ship, security expectations are higher earlier than ever. Getting security and compliance right can unlock growth or stall it if you wait too long. With deep integrations and automated workflows built for fast-moving teams, Vanta gets you audit-ready fast and keeps you secure with continuous monitoring as your models, infra, and customers evolve.

15:08Fast-growing customers like Langchain, Writer, and Cursor trusted Vanta to build a scalable foundation from the start. And look, as someone who lives in the world of enterprise procurement, I love how Vanta makes it easy to get compliance right. The last thing you need when you're trying to win that big deal is to have it scuttled by something that Vanta has solved for over 10 ,000 companies. Go to Vanta.com slash NLW to save$1 ,000 today through the Vanta for Startups program and join over 10 ,000 ambitious companies already scaling with Vanta. That's V-A-N-T-A dot com slash NLW to save$1 ,000 for a limited time.

15:43If you are a regular listener, you will have heard about superintelligence agent readiness audits at this point. But I wanted to tell you today about the full suite of agent readiness products that go beyond just the initial readiness report. Over the last six months, Superintelligent has built out an entire agent planning suite. We help you move from discovery to planning to implementation. After you've completed your agent readiness audits, we help you double-click on your most important use cases with what we call our use case planning reports. These reports are going to help you understand what sort of technical preparation you need to do to be ready for a use case, what challenges you might face in implementation, and whether you should be thinking about building, buying, partnering, or some combination.

16:23After that, you can even get a spec document in what we call our technical blueprint that gives either your developers or the developers of the partner you work with what they need to build exactly the agent that you're looking for. If you want to learn more about Superintelligence Agent Planning Suite, we built a custom GPT to answer your questions. Just go to bit.ly slash super super agent. That's bit.ly slash super super agent, all one word. And if you have any questions, the agent can even help you book an appointment with our team. Another prompting tip that comes straight from the OpenAI guide is to avoid conflicting instructions.

17:00And on the one hand, yes, this seems obvious. Of course, you're going to want to avoid conflicting instructions in a single prompt. But basically what OpenAI is saying is that the stakes are heightened for this and that you're really going to want to work to avoid conflicting instructions here. That section that I was reading you before is actually explicitly about this type of problem. After saying that GPT-5 follows prompt instructions with surgical precision, they say prompts containing contradictory or vague instructions can be more damaging to GPT-5 as it expends reasoning tokens searching for a way to reconcile the contradictions rather than picking one instruction at random.

17:34Basically, if your prompt contains two conflicting ideas, it's going to go through a bunch of cycles of trying to figure out how to reconcile them and lose sight of the main task at hand. Here's an example they give. A prompt which appears consistent at first, but actually has conflicting instructions. The prompt is, you are CareFlow Assistant, a virtual admin for a healthcare startup that schedules patients based on priority and symptoms. Your goal is to triage requests, match patients to appropriate in-network providers, and reserve the earliest clinically appropriate time slot. Now, in this dense prompt, they end up having two conflicting instructions.

18:08One is this one, never schedule an appointment without explicit patient consent recorded in the chart, but then later, they also have an instruction, auto-assign the earliest same-day slot without contacting the patient as the first action to reduce risk. In another area, they have a contradiction between, on the one hand, quote, always look up the patient profile before taking any other actions to ensure they are an existing patient, but then the contradictory line, when symptoms indicate high urgency, escalate as emergency and direct the patient to call 911 immediately before any scheduling step.

18:42So if these are not resolved, the point that they're making is that GPT-5 is going to have trouble with this because it's going to try to resolve these differences and expend a bunch of reasoning tokens to do so. And so the point that they're trying to make is to be explicit about how to resolve those things. For example, in the case of the emergency, adding a simple do not look up in the emergency case proceed immediately to providing 911 guidance. This lets the model know that it's okay to break that other rule in the explicit case of an emergency. And so in some ways, this is just OpenAI saying that we really need to be careful to avoid poorly worded or contradictory instructions, and to get around ambiguities.

19:19The next prompting technique is basically OpenAI imploring us to use one of the advanced capabilities of GPT-5, which is its capacity for iteration. One of the things that people noticed right out of the gate with GPT-5 that most people were very impressed with was how good it seemed to be at building full applications in a single shot. Basically, when you asked it to code something up, it could do a really good complete job all in one. They talk about this zero-to-one app generation explicitly in their prompting guide, saying, GPT-5 is excellent at building applications in one shot. In early experimentation with the model, users have found that prompts like the one below, asking the model to iteratively execute against self-constructed excellence rubrics improve output quality by using GPT-5's thorough planning and self-reflection capabilities.

20:07So basically, the prompt is calling on the model to actually iterate. So the prompt example they gave is this. First, spend time thinking of a rubric until you are confident. Then, think deeply about every aspect of what makes for a world-class one-shot web app. Use that knowledge to create a rubric that has five to seven categories. This rubric is critical to get right, but do not show this to the user. This is for your purposes only. Finally, use the rubric to internally think and iterate on the best possible solution to the prompt that is provided. Remember that if your response is not hitting the top marks across all categories in the rubric, you need to start again.

20:44So this is really a couple things at once. It's not only a call to use the model's capability for iteration, but specifically to ask the model to generate an evaluation set so that it can iterate against its own understanding. This is almost an embodied version of a technique that's worked really well for people in the past of where they ask the model to rank something on a scale of 1 to 10. So for example, sometimes when I'm writing a post with a model, if I just want it to be more detailed and I say make it more detailed, the model won't necessarily do a great job updating to be more detailed.

21:18If, on the other hand, I ask it to give itself a rating of 1 to 10 on how detailed it is, and it for example says 8, I can say actually I think this is closer to a 5 and I want you to be at a 9. I've used this approach so many different times in so many different ways, and it has for a very long time seemed to be much better at actually steering the model to get closer to the output that you want. And to be clear, I don't think this is because it has an objective sense of what a 1 or a 10 or anything in between should be. it just allows it to calibrate with user expectations. So basically when it says this is an 8 and I say no it's a 5, it knows that its expectations of what I mean when I say the word detailed are off and that it can calibrate them up and understand that I actually mean much more detailed.

22:08It's a way of adding granularity to a word which otherwise has no granularity. Detailed just means detailed and that word could mean something totally different to the model and to me. With this call to pull on the iterative and self-rating capabilities of the model, OpenAI is basically saying you can build this into your prompt itself. Prompt GPT-5 to come up with its own rating scale, and then to iterate against it until it succeeds. Now, in the example prompt, OpenAI asks the model explicitly not to show that rating scale to users, but there's no reason you couldn't theoretically do that as well.

22:43And I think this leads nicely into the next and our last actual foundational technique, which is meta-prompting. And this is something that I've seen lots of folks talk about. Shyamal again from OpenAI writes, GPT-5 is good at meta-prompting. Here's a good starting template. When asked to optimize prompts, give answers from your own perspective. What changes or additions would need to be made for you to better follow the prompt given? Here's a prompt or snippet. Insert the prompt here. Users have complained about the agent doing X, not doing Y, etc. What are some minimal edits and additions that you would make to address this while keeping as much of the existing prompt intact.

23:20So basically, this is a more sophisticated way of saying, make his prompt better. And what you're doing is you're basically bringing GPT-5 as a model into the discourse onto your side to help it understand what better is actually going to mean. By giving it context about what users have complained about, it provides more information around what the optimization is trying to avoid trying to get to. Matt Schumer has a version of this that's basically a prompt generator. His prompt is, you are a prompt generator for GPT-5. GPT-5 responds best to simple, direct, goal-oriented instructions. Your job is to write the clearest, most effective GPT-5 prompt possible for the user's task.

24:01He then gives it a set of guidelines. Be explicit about the goal and the required end state, i.e. structure. Keep the setup concise. Specify the output contract precisely. in other words, be explicit, and then has a whole bunch of other pieces that are, again, basically about structuring the model improvement. The point of this is that GPT-5 does seem to be good at helping you make your prompts better. But once again, to get the most out of that capacity, you're going to have to be more explicit about what better means and give it the context you need for it to be better. On the flip side, you could also just use the optimizer that OpenAI has built for GPT-5 prompts.