Too Big an Answer Is a Guess in Disguise
I have written that it does not start with the prompt. Here is the part I left out: there is a proportion. It is not the prompt that decides, it is the ratio between three quantities, and most people count only two. When that proportion is wrong you do not get a worse answer. You get a guess in disguise.
Three quantities, not two
Every AI call has three things in it. The context you put in: your code, your transcripts, your domain data, your own thinking. The prompt: what you ask for. And the output: what you get back. Almost everyone talks about the prompt. Almost no one talks about how the three relate to each other. That is where the answer sits.
The paradox that is not one
I have said that a lazy, short prompt is inexcusable. I have also said the prompt should be small. That sounds like a contradiction until you see that these are two different axes.
A short prompt is inexcusable when it is short out of laziness and the context is thin too. Then there is no direction and no ground. That is vibe. But the prompt should be small relative to a rich context. The variable that should grow is the context, not the prompt. Precision is not the same thing as volume.
Which leads to something worth saying plainly. A long prompt is usually a symptom of too little context. You are trying to compensate for a missing foundation with instructions. You describe a generic scenario in words instead of handing the model your actual material. And then it answers from generic knowledge, the kind that need not be true.
Hold the output against the context
Now take the output and hold it against the context. If you get back more than you put in, the surplus comes from somewhere. It does not come from your truth, because it was not in what you gave. It comes from the model's generic priors. The model simply invents the difference.
This is not the same as the context window being too small, or you only getting back a few thousand words. It is not about how much you CAN get out. It is about where what you get out comes from. Grounded output is a distillation of your context. It is smaller than what you gave, because it boils it down. Invented output is an expansion of a thin foundation.
How they relate
| Context | Prompt | Output vs context | What happens | Verdict |
|---|---|---|---|---|
| Rich, unique | Small, precise | Smaller than context | Distillation of your truth | The ideal |
| Thin | Large | Larger than context | Model fills from the generic | Generic, often untrue |
| Thin | Small and lazy | Anything | No direction, no ground | Inexcusable |
| Rich | Large | Roughly equal | Prompt fights the context | Scattered, over-instructed |
| Rich, unique | Small | Much larger | You asked for more than you gave | Invented despite good ground |
Two things I am not claiming
Size is a proxy. It is not the number of characters that decides but how much unique, relevant material you give. A large window full of junk is worse than a small window with exactly the right thing. When I say rich context I mean dense, not long.
And output larger than context is not always wrong. If you deliberately ask for expansion, turn five bullet points into a draft, you are inviting the generic in. That is right when you are ideating. It is dangerous when you think you are getting facts. The difference is that you know where the surplus comes from.
Why I spend the time on the context
The prompt is the direction. The context is the ground. And the output should never contain more truth than the ground you put in. If it does, you have not got a better answer. You have got a guess in disguise.