Claude and ChatGPT Got More Literal. Your Old Prompts Are Backfiring

0:00 The prompts that worked last year are quietly backfiring on newer models like GPT-5.5 and Opus-4.7. Most people haven't caught it yet. If you're new here, I'm Dylan. I run an AI consultancy, and prompting comes up in almost every single client coaching call that I have. I've been running both

0:17 models against client workflows since they came out. The habits everyone picked up over the last year are now hurting your output, not helping it. I see the same three habits over and over. Let me show you what they are and how to fix them. Let's get into it. There are many things that have changed between

0:34 the previous models and the newest models today like GPT-5.5 and Opus-4.7. But, the primary thing that's changed that's impacting a lot of people's prompts is the fact that these models take you more literally, which means the words that we provide to these models are more important than ever. And the

0:49 fact that these models take our words more seriously is what's impacting these three things I'm going to walk you through. So, here are the three prompt changes we need to make to ensure we're getting the most from these models. I'll walk you through each one of these more detail in a second, but as a quick

1:00 overview, the first thing, which is really hard for me to change, is the fact that you need to stop telling the AI it's an expert in whatever field that it's working on for you. After that, we need to be explicit to the AI if there are specific files it needs to reference to achieve a task. We need to tell it

1:12 what files to look at instead of assuming it's going to look at those files without us stating anything. And finally, with these new models, they can take many steps to achieve really complex tasks now, more steps than they could ever in the past. So, it's important that we ask the AI to tell us

1:25 how many steps it's taken to ensure it's not skipped anything. So, those are the three fixes. We'll start with the first one, which is honestly probably one of the hardest for me to change because I've been doing it for so long. And that's simply dropping the role at the beginning of your prompt. And why does

1:36 this matter? Well, there was a study done recently on GPT-5.5 and Opus-4.7. And what they did is they took the same AI with slightly different prompts on the same task. And all they changed in the prompts were adding a specific role at the beginning with the same task. In the second prompt, they just removed

1:53 the role and had the task alone. The AI itself achieved the task at a much higher accuracy without the role. Now, the theory here is that when you add a role to an AI saying you're an expert writer, an expert strategist, or a salesperson, whatever else, it gets hyper-fixated on that role and spends a

2:08 lot of time thinking about how it's going to convey that persona to the user instead of focusing on the the task at hand. And that's one way that this whole AI focusing on things literally is impacting the quality of the output you're getting from the AI. Quick pause in your regular programming. This video

2:23 is brought to you by me, as always. Two quick things. First off, below is a 30-day AI Insight series, completely free. You'll get 30 insights in your inbox of how 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

2:39 offerings to see if there's a good fit between the two of us. Now, let's get back to the video. And the fix for this is simple. So, this here is a prompt that you would probably give to an older model. But with models like GPT-5.5 and Opus-4.7, you want to change this. Reason being is in this prompt, you're

2:53 spending half of the time asking the AI to think about being a world-class expert in the specific field. And you've only given half of the space for the task, which is help me figure out my pricing strategy. A better prompt here would be something like this, where we remove the role completely and we focus

3:07 more time on what we want and what good looks like. And that matters most for these models now. So, we need to be specific on the goal and the AI will figure out how to get there. So, here instead of saying it's a expert in given field, what we're doing instead is we're saying give me three pricing options.

3:22 Each one should lead with a trade-off and cite the source for any numbers that you used. So, this is very specific on what good looks like and what we want from the AI, not worrying about the role at all. So, this is the first thing that we need to change is removing the role. The next thing we need to change in our

3:36 prompts for these new models is naming the files when relevant. What do I mean by when relevant? Well, if you're doing something inside of a project, so many of you are using GPT projects, cloud projects, Gemini gems, or Copilot agents, they're all the same thing. You have a given AI that's targeted on a

3:52 specific task that you want it to do over and over and over again. Well, in the past with these older models, you could prompt them and not necessarily specify within this project there are context files that it can reference when necessary. Such as your brand voice for writing proposals or past proposals

4:07 you've drafted that it should reference or pricing for those proposals. All of this sits inside of your context files. And the older models would sometimes guess that these files are relevant and reference them. So, you didn't have to be explicit in your prompt. But today, this is completely different because

4:20 remember, these models take you literally. So, if you don't explicitly say that the AI needs to reference these files to achieve a given task, it's not going to reference those files. So, you have to call it out explicitly. I'll show you two examples. So, this here is a before prompt, a prompt that you could

4:35 use in the past where we have an AI writing a proposal for us specifically within a GPT project or a cloud project. And in that cloud project or GPT project, there are series of context files that it should be referencing. In this prompt, we never say anything about those context files. All we do is we ask

4:49 the AI to write a proposal for this specific company based on the discovery call notes that we pasted into the chat but not the context file section. So, here we're hoping that the AI understands that we want it to write a proposal, which means it needs to look at the context files. That would maybe

5:02 work in the past sometimes. But now, since the models take us more literally, we need to use something like this. Where what we're doing is we're telling the AI to write a proposal for this given company. Right after that, we say I need you to look at the context files in this project. Specifically, I need you to

5:16 reference the brand voice file for my tone, the past proposals for what the structure should be, as well as the pricing file to understand any numbers that are relevant for this proposal. So again, we're being explicit on what files it needs to reference so it doesn't skip them. And this may feel

5:30 like somewhat of a downgrade, but it's not because in the past the AI would guess randomly on which files it should look at. And that's not reliable. But now with these models, since they follow instructions so well, that when you ask it to look at the files, the probability that it looks at them is much higher

5:45 than it used to be. We just need to be more explicit and more clear on what we want. So, this is the second change, which is naming the files irrelevant. And the final change of the three changes is demanding an audit or report from the AI after it's achieved a really complex task. And it's important to

5:58 mention that for this specific change, this is more of an emergent problem that we're starting to get because these models have become more effective at complex tasks. Because in the past, with the models that we used to have, they could probably take five to 10 steps for a given task and work between maybe two

6:12 and four minutes. But as these models have become more advanced and more effective at hard to harder tasks, they can achieve tasks that maybe include 50 steps and they could work for 30 minutes to an hour to even six to seven to eight hours. Especially if you're using a tool like Codex or Cloud Co-work on your

6:27 desktop. And the problem that that's arisen here is that in the past, when the AI took maybe five to 10 steps, it was easy for you to audit this and say, "Okay, did it take all the steps?" Because there were only a few. But now, if it takes 50 to 100 steps, it's really hard to audit this unless

6:41 you go through and count every single one. So, we need help with this. And the reason we need help is with these newer models, sometimes what happens is maybe we give the AI 50 items that it needs to process. So, that could be 50 transcripts, 50 presentations, invoices, whatever else.

6:56 And when it processes all of these items, it comes back and says that it's done. But there's a chance that it's actually not. There's a chance that of those 50 items, maybe it skipped eight of them. And you would never know unless you did a human audit on every single one of these items. So, what we can do now is

7:12 we can actually ask the AI to count the items that it processed. And this is the exact prompt I'd recommend copying and pasting into your task that's related to this specific issue. So, in this prompt, what we're doing is we're asking the AI at the end, "After you've finished, I want you to report the exact count of

7:26 items that you processed." Cuz I expected you to process 50, which means I expect you to give me back 50. After this, we ask the AI, "For each item that you skipped, you need to name the item that you skipped as well as tell me specifically why." And then finally, if you do use the word come or

7:42 complete or whatever else, you can only use it when you're actually finished and you've processed all the items. So, in all three of these things we're asking the AI to do, the first one we're simply asking the AI to provide back the count so we can do a quick audit. After that, we're forcing the AI to give us proof of

7:56 when it did skip anything, explicitly giving us the name as well as a specific reason. And then finally, we're setting the incentive of saying you can't say complete or done unless you're actually done. I know this prompt is simple, but it is quite effective. And those are the three changes. So, let's do a quick

8:10 recap on all three. So, the first one, remember you need to drop the role. And again, this is hard for me because I've been doing it for a while, but when you say you were an expert in X, just remove that and focus on what you want the AI to do and what good looks like. That matters way more now than the role.

8:24 After this, we want to name the files. So, if you're using a GPT project or a cloud project where the AI has context files that it can reference, we need to explicitly ask the AI to reference those files. And if there are many of them, we need to specify which files to reference when. Because we can no longer rely on

8:38 the AI to unreliably guess which ones to use, we need to be explicit on those so it can be more reliable on that task. And then finally, we have demanding the report. So, this is more of an emergent property because these AIs can do longer horizon tasks now. So, it can work on 50-step tasks instead of only five to 10

8:53 steps, which means we need to have the AI tell us the exact count of how many steps it's taken because we expected it to process 50 things or 100 things. And also, if it skipped anything, we need it to name exactly what it skipped and why. And that's it. So, as a reminder, two things. First off, below is a 30-day AI

9:09 Insight series completely free. You'll get 30 insights in your inbox of how you can apply AI to your business in 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. So, I want to give you a quick word of caution before

9:22 we end this video. These three prompt fixes will not save your work if it's outgrown the browser in the first place. Some tasks belong in the desktop with agents like Cloud Coworker Codex, and there are seven signs that tell you when it's time to move on. I'll walk you through all seven right here. So, go and

9:39 check that video out. I'll see you next time, internet.