How AI models work

You do not need to understand transformers to write good prompts. You do need four facts, because almost every disappointing answer traces back to one of them.

1. It predicts, it does not look up

A model generates text one piece at a time, each piece chosen as the most plausible continuation of everything before it. That is why it is fluent about things it does not know: an invented policy number is exactly as plausible-looking as a real one.

The fix is not to ask it to be careful. The fix is to give it the real material — attach the file, connect the Brain, paste the clause — so the plausible continuation and the true one are the same sentence.

2. Everything it knows about your request is in the context

The model sees: your message, the conversation so far, the files and pages the task has pulled in, and its instructions. Nothing else. It cannot see your screen, your other tasks, or what you meant but did not type.

That context has a limit. Very long conversations get summarised as they grow — which is fine for the thread of an argument and lossy for exact details. If a specific number matters, restate it rather than assuming it survived from forty messages ago.

3. It has no memory between conversations

A new task starts blank. What persists across tasks in Dhanu is deliberate and visible: the Brain, an AI staff member’s standing instructions, a project’s description, your saved prompts. Anything else you told it yesterday is gone.

This is a feature — it is why a task from last quarter cannot quietly poison today’s answer — but it means recurring context belongs somewhere durable, not in a message.

4. It has a knowledge cutoff, and the internet is a tool

The model was trained up to a point in time. For anything after that — this week’s news, a price, a competitor’s launch — it needs web search, which is a tool it uses, not knowledge it has.

Ask “what changed in the last month?” and a model without search will answer from before the cutoff without noticing. Ask it to check, and it will.

What follows from all four

Symptom	Usual cause
Confident but wrong specifics	No source material in context — fact 1
Forgets a constraint from earlier	Long conversation, detail summarised away — fact 2
Ignores something you established yesterday	New task, no memory — fact 3
Out of date	Cutoff, no search — fact 4

None of those are fixed by asking more politely. They are fixed by changing what is in front of the model — which is what the rest of this section is about.

Next

The parts of a prompt covers what to put in a request so it lands the first time.