(Original en español aquí.)
When you ask something to a human, the human doesn't just receive your request and start working on it without further ado: if someone asks for food because they're hungry, you wouldn't make the same meal for your 4-year-old child as for your grandmother, or for a customer at a super-luxury New York restaurant.
The human receiving the request adds something, some extra information to fill in the gaps that the requester hasn't provided; and they do this with information they can gather from the context, experience, common sense, assumptions, or whatever they can, so the response to that request isn't what the other person told you they want but rather what they really want.
AIs aren't as good at this as humans, so if you let them do this part of the work, the result will be subpar; so you have no choice but to do it yourself. But this means your communication with them is different from what you'd have with a human, and you need to understand how to do it because it's already a change from what we're used to: this is what is called "knowing how to do prompt engineering," which is actually just a snobbish way of saying "explaining yourself well."
So with this, we have the first key point:
To interact with an AI in an optimized way, we need to make optimized requests.
Alright. And how do you do that?
Learning it is easy and you'll see it right away, but to achieve it, you first need to understand how the AI works in terms of workflow. Here are the phases:
1) The AI receives the request. (We're going to learn how to do this well)
2) If it hasn't been told where to get the information it might need that isn't already in the request to solve it, it chooses the sources based on context or the elements at its disposal.
3) It goes to these sources and gathers the information that is available there.
4) It processes the request to generate a result. This process uses two elements:
- The information it now has, whether from the request or obtained externally.
- Processing instructions. If they are in the request and are sufficient, it follows them; if not, it will have to make up for what's missing by deducing from context or available elements, and if that's not possible either, then it will do so randomly.
5) It presents the information to the requester. If the request has sufficient instructions about this, it follows them; if not, it makes up what it can from context or available elements (this usually suffices because there are always default communication rules in its programming), and assuming there's no other way, then randomly (which shouldn't happen).
Have you noticed that in that workflow, there's a lot the AI has to make up as best it can? That's exactly what you do when your 4-year-old tells you they're hungry. That's the part of the information that humans correctly add to a request, and AI doesn't do as well.
And with this, we come to the second key point:
An optimized request is one that already includes all the information the AI needs to work with (except what needs to be sought from external sources), so the AI doesn't have to generate it.
And here comes the solution. How do we make these requests?
Before getting into that, remember the golden rule: what you don't tell it, it has to make up. And if it has to make it up, it will do so as best it can, not according to what interests you. So, the further you deviate from the optimal request, the further it will deviate from the optimal result.
And now we can finally answer the title question. The short version would be "fill in the gaps you don't want the AI to generate," but we can elaborate to help you think about it better and not miss anything:
- Information as complete as possible for processing the request, and correct and precise instructions for finding the correct sources (or, if not possible, at least the correct type of source) from which to obtain the information that needs to be gathered externally.
- Correct and precise instructions to process the information.
- Correct and precise instructions to present the results.
One important thing to note is that sometimes it takes more than one request to get the answer you need. For example, if you're not sure which is the correct source to then order the AI to go and look for what you want, you would first have to make requests about that and then use the results you've obtained to make the final request more optimal. Working with an AI forces you to think more, because you're taking on mental work that would normally be done by the human you're requesting from.
If you look closely, in the workflow we've seen before, there's step 3 where we can't do much (when you go to a place to find something, it's either there or it isn't); but otherwise, we see that the whole process depends on us. The AI follows the instructions to the letter whether they come from the request itself or alternative sources, there's no more to it.
Of course, sometimes what you actually want is for the AI to make things up. For instance, if you feel like chatting, you can just say "hello" and go from there. But clearly, you can control the level of control you exert, depending on what you want the AI to do for you.
So when you're about to interact with an AI, ask yourself if your part of the interaction includes the three aspects I've just mentioned, and if you're unsure about the details, review the previous workflow to make sure you haven't missed anything. For example:
Example of a suboptimal request: "Do you think I have a chance of getting with my desk mate in class?".
Improved request: "I'm an overweight, poor 50-year-old with a hygiene aversion. I've enrolled in a zen decoration course to see if I can score because I haven't gotten lucky in years, and I've been lucky enough to have a supermodel 22-year-old as my desk mate who has a millionaire fitness boyfriend she adores with all her heart. She seems to avoid me and when I try to get close, the first thing she does is cover her nose and start gagging. Based 70% on what I've just told you and 30% modified by what you find in official sources from the Ministry of Social Affairs about couples with different ages and romantic preferences in her demographic group, tell me how many chances I have of her leaving her boyfriend for me. Just give me a number between 0 and 100 as your answer."
The second version is still not optimal, but you can understand why with this one, the AI would likely be more on target... Assuming what the guy wants is reality, of course.