Context, the key to the prompt

Of the five pieces of a prompt, there is one that moves the needle more than the rest. In this step I show you which one it is and why, because the day I learned to look after it my answers stopped coming out generic and started actually serving me.

I come from the step where I took the prompt apart into its pieces. One of them was context, and I named it in passing, as just another piece. Now I give it the place it deserves, because it took me a while to realise how much it weighs.

The AI only sees what is on the table

The image that put everything in order for me was this one: treat every chat as if you were talking to someone very capable but who has just walked in and only sees what is laid out on the table. They don't know you. They don't know what you do for a living, what level you're at, what you need the thing you're asking for, or what happened in yesterday's conversation. The only thing they have is what is written on the screen right now.

All of that, everything in front of it, is called the context: the text the AI can look at to put together its answer. What's on the table, it uses. What you didn't put on the table doesn't exist for it.

A misunderstanding worth knocking down

For a while I spoke to it as if it knew me. I took for granted that it knew who I am, that it remembered what I told it last week, that it understood my situation without my explaining it. And it doesn't. The AI doesn't know you and doesn't keep your chats the way a person does. If something looks like it remembers from an earlier conversation, it's because the system has put it in front of it again, on the table; not because it carries it inside.

This ties back to what you already saw in the first step: behind it there is no one who understands and remembers you, there is a calculation over the text it has in view. And if your case isn't in that text, your case doesn't enter the reckoning.

What you don't tell it, it fills with fog

Here is the practical consequence, the one that really changes how you ask for things. If you want a useful answer for your case, you're the one who has to give it the context. The context you leave out, the AI doesn't guess well: it leaves it generic or makes something up to plug the gap.

Compare them. "Summarise this for me" forces the machine to assume everything —for whom, what for, how long— so it hands you a run-of-the-mill summary, valid for anyone and therefore of little value to you. Whereas "summarise this for someone who knows nothing about the subject, in five lines and with no jargon" puts on the table exactly what was missing, and the answer changes at a stroke. Same text, different brief, because you've given it the context it used to have to invent.

Take another case, clearer still. "Give me ideas for a talk" leaves the machine without the essentials: not what it's about, nor who it's for, nor what for. It'll throw you a list of platitudes that fits any talk in the world. Now say it whole: "give me ideas for a twenty-minute talk on recycling, for ten-year-olds, ending with something they can do at home." The audience and the purpose —who you're speaking to and what you want them to take away— completely change what it suggests. You haven't asked for better, you've told it more; and by telling it more, you leave fewer gaps for the AI to plug on its own.

It's not about dumping everything you know on it, mind. It's about putting on the table what matters for this request: who it's for, what you're after, what limits it has. A good generic answer is worth little; what you want is a good answer for you, and that only comes out if you give it something to work with.

The door left ajar

With this I close the stretch on talking to the AI well. But by pointing at context as the piece that rules most, a bigger question opens up: and how do I manage well everything the AI needs to know, when it's not one sentence but documents, data, whole conversations?

That question isn't answered inside the lone chat. It has a staircase of its own, and that's where we climb from here.

Definitions

- Context: all the text the AI has in front of it at that moment and can use to answer. What you put on the table; what you don't, doesn't exist for it. - Generic answer: the one that fits anyone and therefore doesn't quite fit your case. It shows up when you gave no context and the machine had to assume everything. - Giving context: putting in the prompt, yourself, the information the AI needs about your situation —for whom, what for, with what limits— instead of expecting it to guess.

Further reading

- Mangas Verdes, Qué es el "context window" y por qué afecta a tus tareas diarias con IA — a plain-language explanation in Spanish of why the AI only handles what it has in front of it. https://mangasverdes.es/inteligencia-artificial/es-context-window-por-afecta-tus-tareas-diarias-con-ia_3261_102.html - DigitalPymes, Cómo escribir un prompt eficaz para ChatGPT: el contexto — why good context cuts down generic, unfocused answers. https://digitalpymes.es/como-escribir-un-prompt-eficaz-para-chatgpt-el-contexto/