Your boy had a recent podcast interview, in which he discussed the effects of AI in education. I encourage you to take a listen, not only because it’s a great way to see how I think about AI, but also a way to prove you once and for all I’m not an AI myself, and that there’s a real human behind this newsletter. Here’s the YouTube link.

I’m going to shake your world a little bit today. If someone tells you models aren’t evolving, they are lying to you. AIs are static in some respects, but the ways we use them are changing rapidly, and you’re probably running an outdated prompting guide in your head by now.

That is, what worked a year ago (or even months ago) for prompting models no longer works. Thus, today, we are fixing that by learning:

The days of having to remember your models to do something in a particular way are over. Let’s dive in.

Ready to transform artificial intelligence from a buzzword into your personal revenue generator

HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.

Inside you'll discover:

Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.

Get Your Guide

The first thing I want to make clear is that you must treat every interaction with your generative AI model as a search process.

LLMs are word predictors (or, dare I say, token predictors, which can be an entire word or just a syllable, but I’ll use ‘word’ from now on to simplify things) that predict the next word in a sequence based on the previous words in that sequence.

More specifically, the model predicts the probability distribution over the next word, which is to say: “Based on the text I’ve seen, how would I rank all the words I know as possible next words to that text?“

You probably knew all this already, but I’m more interested in making sure we understand how that ranking takes place.

Models learn by induction, going from the specific to the general.

That is, they learn how words follow each other by seeing those precise words in a particular sequence order multiple times (e.g., if “am” usually comes after “I”, the model is much more likely to predict “am” if it sees “I”).

Therefore, after training on Internet-scale data, the model becomes capable of inferring, with very high accuracy, what word comes next.

But, in reality, the model has just learned to minimize the expected prediction loss across all words on the Internet, which is to say: the model’s responses represent the average response on the Internet.

This is nuanced because mid-training regimes, later stages of training, use very high-quality data, which nudges the model toward better responses. However, it’s still, broadly speaking, a coarse representation of the average quality of the Internet’s data.

The point I’m trying to make is that, although models can very well offer great responses when you search the right way, vague prompting will yield an average response because the model was trained largely on average-quality data. In fact, I’ve talked about this in way more detail here.

Therefore, as I have hinted, good prompting is essential to nudge models in the right direction. But the point here is that I want you to see every prediction the model makes (which is very heavily influenced by what you give it) as a choice.

If you give the model “I like” and the model predicts “trains,” it’s going into the ‘world of trains,’ the area of its knowledge related to them. The same applies to ‘vegetables’, or to any other topic.

For the model, although there are ways we can change that, the choice is always about ‘probability mass’; the model will choose the direction of higher probability, which, in the case of a model that has learned by forming statistical correlations, is always the one that has more continuations… even if that doesn’t necessarily imply overall sequence quality (in fact, a very relevant criticism against LLMs is that they don’t have the meta-cognition ability to ask themselves if they are going in the right direction).

Circling back to the example above, ‘vegetables’ is more likely to get chosen by the model because, during training, it saw more sequences with that particular choice of words than with ‘trains’.

However, and this is the crux of it all, higher probability mass doesn’t mean greater quality; quite the contrary, as the model gravitates toward areas it has seen more of, which means it moves by default into areas that are “more common”.

So, what can we do? As a general mindset toward model interactions, you should always:

All tips below fall into one of these two categories. But what makes a prompt good for a model?

Just like your choice of words influences how people respond to you, models behave the same way. And the first step is that, well, sometimes it’s worth it to just try again.

For all the talk about clever prompting, the thing that will reliably move the needle is still boring iteration. You try something, it half‑works, you scrap it, you try again.

Most of the time, the difference between a casual user and a power user isn’t that the power user gets it right on the first attempt; it’s that they’re deliberately using each attempt as a micro‑experiment.

A useful way to think about it is: every prompt you send is a hypothesis about how the model’s internal search will unfold. If the answer misses the mark, you don’t just “try again with more detail,” you ask what failed.

I always say that if someone read some of the “stupid” questions I ask models privately, they wouldn’t believe I make a living giving my opinion on hard stuff because of my questions being “stupid”. But that’s the point, sometimes you need to appear stupid and ask the most basic questions to truly capture the essence.

In some cases, it may appear as if the model is the stupid one. Did it misunderstand the goal? Ignore a constraint? Get lost in old chat history?

But probably it’s that your prompt isn’t good enough. That’s the beautiful nature of retries; the diagnosis from the previous try informs the following prompt. Eventually, over a handful of rounds, you’re probably going to get a great answer from the model.

Furthermore, this introduces the first point on context management: context rot is the silent enemy of this process.

Long chats rife with half‑baked instructions, wrong assumptions from the model, and the lot aren’t only going to make things harder for the model at every turn, but they also lead to context rotting.

That’s where “plan → execute → compress” becomes a convenient habit. Just scrap the conversation and write a more or less detailed summary, then start a new one.

The entire point is that you’re continuously squeezing messy interaction history into a small, fresh context that the model can reason over cleanly. It feels a bit like refactoring a messy codebase as you go, and it keeps your iterations fast instead of degenerating into one giant, haunted chat.

In other cases, the model may have gone too far into a given search direction that the conversation is no longer useful at all, and you’ll have to scrap it entirely (i.e., don’t even try to nudge the model back; it’s like trying to make a U-turn at max speed with a Ferrari). Sometimes, the mistake is so apparent that you might consider ditching it after a single turn.

At this point, you may be tempted to believe ‘putting effort into your prompt’ is just dumping as much context as possible.

But funnily enough, getting too detailed with your prompts is actually one of the worst things you can do, no matter what your typical LinkedIn influencer might tell you.

As I was saying, just dumping all context to nudge the model in the right direction won’t work. This was the old view of “prompt engineering”: you mostly control a short system prompt and a user message, so the shortest solution was to dump it all.

If that was your approach, you’re living in the past. Let’s fix that, and let’s make our AIs “smart” and autonomous.