LLM 101 — 01: Prompt as Conditioning
A prompt doesn't command a model — it conditions a probability distribution, reshaping which continuations become likely.
There is no shortage of advice on how to prompt an LLM.
Be specific. Give it context. Assign a role. State the constraints. Show examples. Do not over-constrain it. Tell it how to think. Or perhaps do not tell it how to think.
A surprising amount of discussion has accumulated around the question of what a good prompt should look like.
But what is an LLM? And what, exactly, is a prompt?
In one sense, an LLM is just an absurdly complicated conditional probability distribution.
Very roughly,
Given some context , there is a distribution over what may come next.
There is obviously a lot hidden inside that notation. But for the moment, the complexity is not the interesting part.
The interesting part is the vertical bar.
The model generates something given something else. And a prompt lives on the right side of that bar.
A prompt is a condition.
Conditioning is a rather ordinary idea.
Suppose I ask for the probability of rain tomorrow. There is some distribution of possibilities. Then I tell you that a hurricane is approaching the coast.
The weather did not change because I spoke. What changed was the information we were conditioning on. The same possible tomorrow now carries a very different distribution of probabilities.
Language works this way too. Take a word like bank. On its own, several meanings are plausible. Put it next to river, and one interpretation becomes natural. Put it next to mortgage, and another one does.
The context did not rewrite the definition of bank. It changed which meaning made sense here.
This is close to the intuition I want for prompting. A prompt does not need to create a behavior in order to make that behavior much more likely. It changes the context under which the model generates.
This is slightly different from the picture suggested by the chat interface.
The interface looks like a command line: we write something, and the model does something. So it is natural to imagine:
The prompt specifies an operation; the model performs it.
That picture is useful. But it hides something important.
An instruction is only one thing that can appear in . So is a question, a role, a piece of source text, a formatting constraint, an example, the whole previous conversation. They may mean very different things to us. We give them names like instruction, context, role, constraint, and demonstration.
To the process of generation, however, they all become part of what the model is generating under.
So perhaps the more useful picture is not:
but:
The prompt does not point to one answer. It changes which answers become plausible.
Consider two questions.
Why do we keep returning to a place even after we know it will disappoint us?
And:
What explains repeated return to an environment despite a known negative outcome?
They are asking, approximately, the same thing.
Neither prompt says:
Write like a novelist.
Neither says:
Analyze this as a behavioral scientist.
And yet they do not feel like the same question. The first seems to invite ideas like memory, attachment, expectation, familiarity, perhaps even home. The second makes things like persistence, reinforcement, learned behavior, incentives, or expectation updating feel more natural.
The difference is not merely that one answer might sound more literary and the other more technical. Something deeper has shifted: what counts as an interesting explanation, which vocabulary is natural, which distinctions are useful, what level of abstraction feels right, even which analogies seem appropriate.
The prompts contain almost the same proposition. But they do not place the model in the same place.
That phrase — the same place — deserves to be taken seriously.
Answers come in kinds. A vocabulary, a level of abstraction, a sense of what counts as relevant, a style of explanation — these properties travel together, forming coherent regions of possibility. Memory, attachment, home live in one region. Reinforcement, incentives, expectation updating live in another. The two prompts did not change the model, and they did not merely change the odds of some individual words. They moved probability mass from one region to another.
There is no literal map of these regions inside the model. But this is the useful picture of what conditioning does: not selecting an answer, not nudging a word here and there — redistributing plausibility across regions of semantic possibility.
Prompting is conditioning.
And this changes what prompt design is about.
Seen from the interface, prompting is wordsmithing: search for the phrasing that makes the model understand you. From there, the natural questions are all about words — which phrase, which order, which tone.
Seen from the distribution, the words are only the mechanism. What you are actually choosing is a destination: which region of the model’s possible behavior the probability mass should settle on. Two prompts with very different wording can be the same prompt, if they condition the model into the same region. Two prompts that differ by a single word can be entirely different prompts, if they do not.
Prompt design is distribution design.
One question is left open.
Everything in a prompt conditions the model. But not everything conditions it in the same way. An instruction — be concise — contains what you chose to say, and nothing more. A worked example is also context, also on the right side of the bar — but it is a complete, fully specific answer, with every one of its properties already in place.
Both move probability mass. They package information in completely different ways.
That difference is where the next question begins.