Getting Real Work out of a Model · Chapter 01
What is a prompt, really?
4:34 · Narrated with a synthetic voice. Captions are available in the player.
Read the transcript
What is a prompt, really? Not the box on the screen. When you press send, the product assembles one flat block of text: a system prompt you did not write, the whole conversation so far, and your new message at the end. The model reads that block, answers, and forgets. There is no memory behind the curtain. When you send message twelve, messages one to eleven are re-sent as text. That is why a mistake in turn three is still in the room at turn thirty. So the prompt is not your last line. It is everything the product packs into that block, and the work of prompting is deciding what belongs there, what to cut, and what must be said again because the model will not remember it on its own. Name what enters the block before you argue about wording. Send the same prompt twice and you get two different answers. That is not the model changing its mind. At every step it produces a probability for every possible next token, and something outside the model picks one. That picker is the sampler. Its main setting is temperature. Keep temperature low where there is a right answer; raise it when you want range. Two things follow. Pressing regenerate draws again from the same probabilities. And a prompt that only works on a lucky draw is not a prompt you can trust for work. Reliability means the same instruction, under a fixed sampler, keeps landing in the band you need, not once in five tries. A working prompt fills six slots. The task: a verb, an object, and a finished state. The context: the material itself, plus the facts that change what a good answer is. Examples, where a rule is easier to show than to state. Constraints, few and checkable. The output shape, including what to leave out. And the escape hatch, which says what to do when the task cannot be done as asked. Notice what is not among them: a long speech about being helpful, or a threat that the model will be punished. Those lines burn tokens and rarely change the answer. The six slots do, because they tell the model what finished looks like. People often polish the instruction and starve the material. If you ask for a summary without pasting the note, the model invents a note that sounds right. If you ask for a reply to a customer and hide the complaint, you get a generic apology. The context slot is not decoration. It is the work. Put the document, the numbers, the names, and the constraint that matters in the block. Then ask for one finished thing. A short task on rich material beats a long task on empty air. That is the first habit this course builds, before style tips and clever openers. Even a good prompt can be wrong in a confident voice. So you need a check outside the chat. Read the answer against the source you pasted. Ask for the quote or the row that supports a number. If the product can call a tool, prefer the tool for facts. If it cannot, you are the tool. The module’s lessons walk the six slots in detail, with checks and counter-examples. Leave with one question you can use at work tomorrow: before you ship this reply, what did you open that the model did not invent, and would a colleague find the same support? A prompt is the whole block the model reads, not the text box. The sampler explains why the same words can land twice. Six slots carry real work; fluff does not. Put the material in, ask for one finished thing, and check outside the chat. The next modules make each slot practical, from examples to output shape. For now, take one change into your next message: what will you put in the block that is not already in your head, and what will you refuse to ship until you have checked it?