So if you've been using the latest AI models, maybe chat GPT -5, you might have noticed something a bit weird. Yeah, like you know you're using this incredibly powerful tool, right? State of the art. Exactly. But the results, they can be all over the place. Sometimes amazing, sometimes not so much. Right. You get this brilliant piece of analysis one minute, and then next the answer feels kind of lazy, almost worse than the older
model sometimes. It's confusing, isn't it? You're paying for the premium model expecting top quality every time, but it dips. So why is this super smart AI acting well? Kind of inconsistent. It's a fair question, and it feels backwards. But the issue isn't really the AI model itself. It is powerful. OK. The problem comes from a couple of big changes OpenAI made in how the whole system worked. Basically, the way we used to write prompts, it's tripping up this new setup. Oh, OK. So our
old habits are the problem. Let's dig into that, then. Yeah. The goal today is really to give you, the listener, a clear strategy to get back to those consistently high quality results. Definitely. We're going to unpack these two main changes. First, this invisible router thing, and also how the AI now follows instructions very strictly. And then we'll jump into five specific tips. We'll start simple with easy nudges and build up to something we call the perfection loop.
It's a bit more effort, but wow, the results. All right, let's get into it. Segment one. Yeah. Why the old ways are failing. This invisible router. What's that about? I remember you used to be able to pick different models. You did, yeah. You could choose GPT -5, maybe a thinking version. You felt like you had control. Right. Now the interface seems simpler. But behind the curtain, when you send your prompt, it hits this invisible router first. Think of it like an automatic
traffic cop for your request. A traffic cop? But if I'm paying for the best model, why does it need routing? Shouldn't it just go to the best one? That's the logical thought, yeah. But here's the catch, and where the user frustration often comes in, those really powerful models, they take more computer power, they cost more for OpenAI to run. Ah, so the router is maybe
optimizing for cost. You got it. It's often engineered to find the most efficient path, which usually means the fastest and cheapest AI model available that might be able to answer the query. So it's defaulting to good enough to save resources, even if I wanted great. Precisely. And that good enough model gives you those, well, less impressive inconsistent answers sometimes. That's the hidden mechanical reason for the quality dips. OK, that makes a lot of sense, actually. So that's the
router. What about the other big shift, the training? Right, the second piece is how it's trained. Chat GPT -5 got a lot of training focused on serving AI agents. These are programs, basically, that need instructions followed perfectly. No mistakes, no guessing. OK, so it got really good at following orders exactly. Super good. But the flip side for a regular user like you or me, if your prompt is a bit vague or you leave things out, It's much worse now at trying to
figure out what you meant. It won't fill in the gaps like the older ones might have. Nope. It sticks rigidly to what you typed. No implied instructions. That sounds demanding. It can be. And honestly, I still mess this up sometimes. You know, I rush a prompt, make it too simple. Happens to the best of us. Right. Just last week, I asked for a quick summary and it gave me this long definition of terms I obviously already knew. Slight chuckle. Took me a few tries to
get it right. It catches you out if you're not precise. OK, so the router pushes towards cheapness and the AI demands exact instructions. So the big question then, how do we make the router pick the smart expensive model when we need it. Ah, well, that's where our specific phrasing comes in. We have to signal it. Leading us nicely into the quick wins. Tip one. Tip one. Use router nudge phrases. This is probably the easiest thing
you can do. Seriously low effort. OK. You just add a specific little phrase right at the end of your prompt. It's like a little flag for the router that says, hey, stop. This one needs actual thinking power. What are these magic words then? What tells the router to wake up the big guns? The ones that work consistently are phrases like, think carefully about this, or think deeply about this. Sometimes think hard about this works,
too. Hmm. OK. Simple enough. Yeah. Using one of those basically forces the router to send your request to a more capable model, avoiding that quick, cheap default route we talked about. I can see that being useful. Like if I'm asking it to, say, outline pros and cons for a big decision, adding, think deeply about this, would probably get me a much better, more nuanced answer than just asking the question straight. Exactly. You should definitely use it for anything important.
Business plans, you know, data analysis, drafting a critical email, things where quality really matters. Oh, and a quick heads up, don't use emotional language. Saying things like, this is really important to me doesn't seem to work. Right. It's a machine. It needs commands, not feelings. Precisely. Keep it clear. Keep it command -like. OK, so that's nudging the router. What about the output itself? Tip 2 is about length,
right? Yeah, controlling the output length. Because the router also kind of defaults to shorter answers, again, to save processing time. So we need to be explicit about how much text we actually want. Makes sense. So for short stuff, like just the key points. Right. For a short output, you'd say something like, summarize the main points
in under 100 words. Perfect for maybe a quick project update email or you know a tweet Okay, and if I need a bit more like the main points plus some background That's your medium output try asking it to explain this topic in about three to five short paragraphs That works well for explaining something like why your websites click through rate dropped Maybe to your team gives context and then the big stuff full reports articles detailed full output be specific provide
a detailed and full analysis around 600, 800 words, or even more, like write a comprehensive guide of about 1 ,200 words. That gets you the full document. Numbers seem key here. Specific word counts or paragraph counts. Absolutely. And a little pro tip here, use a text expander app. You can save these length commands like summarize under 100 words and just type a short code like S100 to insert the whole phrase. Saves
a ton of time. Nice. Okay, so we can nudge the router and control the length, but what about improving the actual prompt before we send it? How do we make the prompt itself better? Great question. That leads us to using a meta prompt. We basically get the AI to help us write a better prompt for itself. Okay, tip three, the meta prompt optimizer. You're saying we can use the AI to improve our instructions for the AI. Exactly. It sounds a bit circular, maybe, but it works
incredibly well. Most people don't know OpenAI has internal tools for this, but we can kind of replicate it with a specific type of prompt. A meta prompt. Yeah. Which is just a prompt about prompting. Yep. You're telling the AI to act like a prompt expert and fix your request. So wait, we're asking the AI to critique our prompting skills. Is that really necessary for, like, just asking for a blog post idea? Well, maybe not for every single query. But remember how strict
it is now? Write the exact instruction following. By having it analyze your prompt first, you force it to clarify everything. You guarantee the instructions it finally acts on are super clear and detailed. The command is usually something like, you are an expert prompt engineer, analyze the weak points of my original prompt below, and then rewrite it to be much clearer and more effective. Hmm, okay. I can see how that would force clarity. Do you have an example? Sure, think about a vague
request. Write a blog post about working from home effectively. Pretty standard prompt. Right. But the meta -prompt process might turn that into something much richer. The rewritten prompt would likely define the AI's role, like you are a productivity expert. Specify the target audience, remote workers new to the concept. Outline the required structure. Intro, three main tips with examples, conclusion, and set the tone, encouraging and practical. Wow, okay. That's way more specific.
The difference in output quality must be huge. Night and day. That detailed prompt just gives the AI so much more to work with accurately. And if you start doing this meta -prompting regularly, I guess your prompts naturally start looking more structured. Which brings us to tip four. the XML sandwich. Exactly. You'll notice those improved prompts often use these angle brackets like this tag or context. These look like XML tags. OpenAI actually recommends this structure
internally. It helps organize complex instructions. I like the analogy you used before thinking of them like clearly labeled boxes. Instead of just a messy paragraph of instructions, you're giving it distinct blocks of information. Audience, tone, goal. Precisely. It breaks it down logically. The AI processes structured information really well. Let's take an example. maybe asking for a study plan. Right. A bad prompt is just, make me an IELTS study plan, I need band 7. Vague.
A good structured prompt would use tags. Current level band 6, current level, target score band, 7 .5 target score, week area speaking fluency, complex grammar week areas, study time 10 hours week study time. Much clearer. Each piece of info is neatly packaged. And the result you get
back reflects that clarity. Instead of generic advice, you'll likely get a detailed table, maybe in markdown format, with specific activities scheduled like practice part two cue cards, 30 minutes, or shadow native speaker audio, 20 minutes. It becomes actionable. Yeah, that structure is powerful, especially I imagine if you're setting up custom instructions in chat GPT or building your own custom GPTs. Definitely essential for
those. OK, so we've covered nudging, length, optimizing the prompt itself with meta prompts and structuring it with XML. This is all about crafting the input. But how do we kind of guarantee the output quality before we even see it? Ah, that's the final piece. We embed quality control inside the prompt itself. We make the AI check its own work. Interesting. Let's take a quick break and then dive into that final tip. Sounds good. Mid -roll sponsor read. We'll be right
back after the break. All right, we're back. We were just talking about making the AI guarantee its own quality. What's the final tip? Tip five, the perfection loop. Now this one takes a bit more thought when you're writing the prompt, but the payoff is potentially huge. Okay, the perfection loop. How does it work? Instead of just asking for the output and hoping it's good, you first instruct the AI to define what a perfect result would even look like for this specific
request. So you make it create its own checklist first, based on my goals. That's clever. Exactly. You tell it, okay, first... Figure out the criteria for a perfect answer here. Maybe that's uniqueness, clarity, engaging tone, fits the brand voice, whatever is relevant. Then you tell it, draft an answer. Then use your own checklist to grade that draft. Keep refining it internally until it scores a 10 out of 10. Only then show me the
final result. Whoa. So it does the drafting, the critiquing, the editing all internally before I even see anything. Yep. You essentially push the quality control step. back onto the AI before it delivers. You only get the polished version. That's kind of amazing. So, for example, if I needed a content strategy. You could tell it. First, define the key elements of a successful Q4 content strategy for a B2B SaaS company. Make
a checklist. Then, generate the strategy, review it against your checklist, revise it until it's perfect, and then show me the final 1010 strategy. It assesses itself on things like scalability, tone, alignment with goals. Or, for say, a YouTube script. Create a checklist for an engaging tutorial script. Hook, clear steps, visual cues, call to action, friendly tone. Write the script, grade it, rewrite it until perfect, then output. Wow,
just thinking about that. Imagine scaling that kind of internal self -correction across, I don't know, millions of tasks. It's not just prompting anymore, that's like... managing an automated quality process. It really shifts the dynamic, doesn't it? It's best for those really complex, high -stakes tasks. You know, writing entire business plans, generating code that needs to work first time, crafting really long, important documents. Okay, so that's the fifth tip, the
perfection loop. Now, the big takeaway here seems to be that these tips aren't really isolated tricks. Not at all. The real power comes when you start combining them. You layer them together to create what we sometimes call a super prompt. Right. Can we walk through building one, maybe for that project proposal example you mentioned, a fashion retail app? Sure. Perfect example. So first we'd start with structure. Tip four,
we use those XML style tags. Okay, like company context, dot company context, problem statement, problem statement, and then maybe tags for each required section of the proposal. Section one introduction, section two solution, et cetera, laying it all out clearly. Exactly. Then we add tip two. Length control. We specify this proposal should be detailed and complete around 5 ,000, 1 ,200 words. Give it a clear target. Got it. Structure, then length. What's next? Now we bring
in the big one. Tip five, the perfection loop. We'd add instructions like, before writing, define an internal checklist for what makes a project proposal highly persuasive and likely to be approved. Grade your draft against this checklist. Refine it until it achieves a 10 -10 internal score, then present the final version. Okay, so structure, length, internal quality control. Are we missing anything? Just the final touch. Tip one, the
router nudge. Right at the very end of the entire prompt we add, think very carefully about this. Ah, the signal to use the powerful model. Yep, so you see how it all stacks together. Structure provides clarity, length sets boundaries, the perfection loop ensures quality, and the nudge phrase makes sure the right brain is doing the work. That combination feels really robust. It addresses both the router issue and the strict instruction following. It gives the AI everything
it needs to succeed based on its new rules. OK, so wrapping this all up, what's the main thing people should take away from this deep dive? I think the core idea is pretty straightforward. The era of just typing vague one -sentence prompts and hoping for the best. That's kind of over, at least for consistently high quality with models like GPT -5. Yes, the AI is more powerful. But that power comes with these new conditions, the invisible router, the need for absolute clarity.
So our approach has to adapt. It's less about needing fancier tech and more about us being better communicators. More structured, now we ask. Exactly. Clear instructions, better organization. You can start small, you know, just add those router nudges and specify the length. That alone makes a difference. Right. Easy wins first. Then as you get comfortable, start playing with the meta prompts, use the XML structure for complex requests, and try out that perfection loop for
really crucial tasks. It feels like these techniques aren't just about getting a better answer from the AI. They're also about building better processes. for working with the AI. That's a great way to put it. It's about setting up an effective partnership. You define the standards very clearly, and the AI has the power to meet them precisely. So here's a final thought to leave our listeners with, something to mull over. We've talked about how this AI is trained to follow exact commands perfectly
now. Super literally. What do you think happens if you give it two instructions that are perfectly clear, perfectly structured? but they directly conflict with each other. What does forcing that kind of logical paradox make the system prioritize? Something to maybe experiment with. Ooh, that's a fascinating question. What breaks first? Definitely something to try. But for now, maybe just try combining those first couple of tips, the nudge and the length control, see what happens. Good
starting point. Thanks for breaking all this down. My pleasure. Always fun talking about this stuff.