Mega-Prompts Are Dead on GPT-5.5. Here's What Replaces Them.

Added: 2026-05-08

[00:00:00]GPT5.5 came out just last week and the prompts that work 6 months ago are quietly breaking inside of it now.

[00:00:07]New models think differently. So those long step-by-step prompts are actually getting in their way. If you're new here, I'm Dylan. I run an AI consultancy and I spent the last 2 weeks rebuilding prompts with my coaching clients for this exact model.

[00:00:21]So I'm going to walk you through the four-part structure that replaces the old way of prompting and I'll show you why it works better now. So let's get into it. First things first, this model change isn't just only GPT5.5.

[00:00:35]It applies to Opus 4.7, Gemini 3.1 Pro, really any of the state-of-the-art models. The thing about these really intelligent models is they often know how to get to a given destination better than you and I do, which is harsh but it's true and it's only becoming more true day by day. So in the past, some of the best practices associated to prompting meant you had to have a really detailed prompt that had really clear instructions on what the AI had to do step by step. So it looked something like on the left where you would have the path and you would take control. You would say you do step one, step two, step three, etc. Then you finally get the output. And this still is true today for some use cases but that you number of use cases a percentage is decreasing rapidly. And what's more true today for the cutting edge models like GPT5.5 and Opus 4.7 is that all we need to do is we need to be clear on the destination. Where are we headed? And if we're clear on that, the AI will figure out the most effective path to get to that destination. And that's what we're going to talk about today. How do we update our prompts for these new models so we can get the most out of their intelligence cuz right now, if your prompts are very detailed and have a ton of steps, you're actually getting in the way and you're bottling the AI's intelligence. You're holding it back. We need to gradually get out of the AI's way so we can get more and more intelligence out of it over the weeks, months and years to come. And this adjustment in prompts is just the first step. Now what are these adjustments?

[00:01:50]Well, there are four D's and these are the things you need to consider when updating your prompts today and creating new ones in the future. I'll do a quick overview here but we'll go in detail on each later on.

[00:01:59]So the first D is destination. So we've already talked about this. We need to be a clear in our prompt of where we're headed. After that, we have definition.

[00:02:07]So then we need to specify to the AI what good looks like. After that, we have doubt. So this one's interesting.

[00:02:12]As the AI gets more intelligent, it also gets more effective at lying to us, which is not great. So it can fabricate fake data and really convince you that it's true.

[00:02:22]So we need to ensure that every prompt that we write for use cases that matter to us, we need to have proof associated to the claims and the facts that the AI provides.

[00:02:29]And then finally, the fourth D is done.

[00:02:32]Now this one's also interesting as well is when you're using a high-end model like GPT5.5 and you set the reasoning to extra high or heavy, the AI could think for a really long time unnecessarily. So it's both wasting your time as well as your money because that costs tokens. So for many use cases, we need to set a finish line. We need to tell the AI when it's done, finish and then give me back the output. And these are the four D's we're going to walk through. We'll start with the first one which is destination.

[00:02:55]Quick pause in regular programming. This video is brought to you by me as always.

[00:02:59]Two things. First off, below is a 30-day AI insight series completely free.

[00:03:04]You'll get 30 insights in your inbox so you can apply AI to your business and your work. The second thing is if you'd like to work with me, below are a series of offerings to see if there's a good fit between the two of us. Now, let's get back to the video. And we've already talked about this a bit where we're specifying to the AI where we're headed and it determines the path. So I'm going to show you some bad examples of what we used to do in the past and what good examples look like today.

[00:03:25]So that example in a prompt is simply telling the AI to summarize a meeting transcript.

[00:03:30]A better example of this is turn this transcript into a follow-up email I can send to a client. The reason this is better is we're telling the AI not just to do a task but we're telling it why we're doing the task. We want you to look at the transcript so then you can send a follow-up email to a client. So we're being more specific. Another example of good and bad.

[00:03:49]A bad example is make a table from this spreadsheet. A much better example is find the three problems in this spreadsheet that would change my decision for fill in the X criteria. Why this is better is we're not focused on the output format alone. We're telling the AI our intent. We're saying the reason you're looking at the spreadsheet is that I need to understand the three problems that are going to impact this decision I'm trying to make on this data set. And the AI will then determine the best way to then convey that to you so you can make that decision effectively and quickly. So here we're specifying the destination not just how to do it.

[00:04:22]The next D is definition. So this is all about specifying what good looks like because often times we have our own criteria of what matters to us.

[00:04:30]So just some examples of what good could look like is having a very specific format associated to your brand. So being on brand for the company for a presentation, a proposal, a dashboard, whatever else. Another one is including verifiable claims so you can quickly audit the output of the AI to ensure that it's correct. This is just a small sampling of some criteria of what good looks like. You probably have tons of ideas of what good looks like to you and your use cases. And let me show you what a good prompt looks like in defining success criteria for an AI. So here in this prompt, we're stating that we want the AI to rewrite this email so it's clear, calm and direct. Keep the same facts in the email I've given you as a draft and then keep it under 200 words as well as put the ask in the first three sentences.

[00:05:09]Now I put stars next to these two sentences because for me, these are much better than the first two.

[00:05:16]Now why is this better than the first two? Well, these two points are binary.

[00:05:20]So if you can, in all cases, try to make the success criteria of what good looks like binary instead of on a spectrum.

[00:05:26]These are binary because I can easily understand if something is under 200 words or not. It's a yes-no answer. And the same thing applies to the next sentence, which is putting the ask in the first three sentences. If it's not in the first three sentences, it's wrong. Now the reason we want binary criteria is it's easier not just for you to validate but also for the AI to check its own work before it gives you an output. So it can get closer to good much faster with binary criteria. So that's our second D, which is defining good.

[00:05:52]Our third D is doubt. So there were some benchmarks and tests done with GPT5.5 and many of the other state-of-the-art models recently and they found that these models were right more often but also they guessed more often as well from previous models. And not only did they guess but they did it more confidently and convincingly. And that's where this third D comes into play.

[00:06:10]It's okay for us to rely on the AI to determine the path of getting to a destination but it's not okay for most use cases where liability matters such as financial liability, legal liability, brand reputation, things like that. It's not okay for us to rely on the AI solely in those cases. So anytime that the AI gives you facts or claims on the output, you always want to specify where it came from. It needs to provide proof. So it gives per claim, it gives you a source in brackets and you can easily audit where it came from. And that's important here because you don't have to read through this and try to grok what happened and where it came from. You want it to give you proof so you can quickly spot-check the source versus what it stated. Now let's look at some examples of what good and bad look like here.

[00:06:49]So in the past, a lot of people, they would just say don't make stuff up.

[00:06:53]That's a really bad way of doing this. A much better way is saying something along the lines of after every fact or claim, I want you to cite the source in line looking something like this where it says source colon and then gives a specific report and or page number associated to that that claim or fact.

[00:07:08]That's one example. Another example of good and bad is the bad side is asking the AI to don't hallucinate. This will likely do nothing for the AI's performance. A much better version of this is when you're not sure, I want you to write unverified or just leave it blank. I'd rather see a gap than a guess. Now what we're doing here in this prompt is we're changing the incentives of the AI. Because the AI's sole purpose in life is to make you happy and the way that it's going to do that is by giving you answers. So if it can't find the answer to a question, it's going to fabricate it to then make you happy. So we need to let the AI know that it's okay for you to give me blank answers than wrong answers and that's what we're doing in this prompt. So this is our third D, which is doubt. Our fourth and final D of improving our prompts is done. Now as a reminder, the reason we're doing this now is that these really high-end intelligent models such as GPT5.5 with extra high reasoning or heavy reasoning enabled, they can work for hours, which is useful in some rare use cases. But for a lot of the use cases you're using AI for, that's not relevant. It's also a waste of time for you and also a waste of money for tokens. And what we want to do is we want to set a finish line for the AI in the prompt. So once it reaches that, it stops. And there are many ways you can do this. I'm just going to show you two examples of what good and bad look like.

[00:08:18]So the old habit of what a lot of people would do is they would tell the AI, be exhaustive, think deeply and research everything. We want it to be super holistic. A better version now for these high-end models, especially if you're using the extra reasoning so extra high inside of Codex for Open AI or heavy inside of ChatGPT, you want to set something up like this.

[00:08:36]Stop once you can answer the main question with enough evidence.

[00:08:39]So we're telling it when to stop specifically in relation to the task that we gave it.

[00:08:44]That's a good example. Another good bad pairing is a bad example would be to cover every angle. We want it to be holistic. And this is no longer necessary for this type of prompt, especially with these models.

[00:08:55]So instead, what you can say is when the output meets the checklist, give me the final version.

[00:09:01]Assuming you provided the AI a checklist of what you want back, it can look at that. Once it's completed it, it can then send you the final output. And this is our fourth and final D of improving our prompts, which is done.

[00:09:12]Now I want to show you a side-by-side of what this looks like completed. So on the left side is the old school prompting that many of your prompts probably still look like, which is a mega prompt with many instructions. So in this example, we're telling the AI, I want you to act as a world-class business strategist. First, read the following transcripts. Then identify the key themes. Then extract action items.

[00:09:30]Then write the client email, etc., etc., etc. We're giving it all the steps it needs to do. This is no longer necessary.

[00:09:37]Instead, we want to have a prompt that looks something like this where at first we're telling the AI where the destination is. So we're saying I want you to turn this transcript into a client-ready follow-up email. After we've given the destination, we then move to definition. So we're telling it what good looks like. And here we're saying success means the email clearly states what we decided, what is still open and the next action for each person. After the success criteria, we then move on to doubt because again the AI can hallucinate to us so we want proof. Here we're stating use only the decisions directly supported by the transcript. Put unclear items under the open questions. These two sentences are doing two things. This first sentence here is grounding the AI. So we're grounding the AI in the document ensuring that it isn't pulling information from its head or the internet, only the document. The second sentence, we're telling the AI that it's okay for you to give me blank answers. I don't want wrong answers. And then finally at the end of this prompt, we're telling the AI where to stop. So here we're saying when the checklist is met, give me the final email. Assuming that you provided the AI checklist to then check off in the process. Now this is what it looks like all put together.

[00:10:37]Let's do a quick recap of the four D's.

[00:10:39]Now first off, the reason we're doing this is that new models like GPT 5.5 and Opus 4.7 are extremely effective at knowing how to get to a specific destination. We just need to tell it where to go and it'll figure out the rest. And we'll update our prompts with these four D's for the use cases that matter most to us. Again, first is the destination. So we're telling the AI where we're headed. After that, we're being very specific on what good looks like. And here if possible, you want to use binary criteria, yes no things.

[00:11:04]After this, we're going to make sure that for the use cases where it matters, financial, legal, and brand reputation, we want to have proof associated to each fact and claim the AI gives us because again these models are very good at lying. And then finally, if you're using extra high reasoning or heavy reasoning for ChatGPT or Codex, you don't want it to run forever wasting time and tokens.

[00:11:22]Instead, you want to tell it exactly when to stop based on the finish line you provided in your prompt. And that's it. So as a reminder, two things. First off, below is a 30-day AI insight series, completely free. You'll get 30 insights in your inbox so I can apply AI to your business and your work. The second thing is if you'd like to work with me, below are a series of offerings to see if there's a good fit between the two of us. Now after that conversation, I want you to picture this. You write a great prompt using the four D's we just discussed. It works so you save it. But where? A project? A skill? I watch my clients freeze at this exact step every week. And I made a video on how to choose right here. So go ahead and click it. Go ahead.

[00:12:01]I'll see you next time internet.