Prompt
The written request you hand an AI, spelling out what to do
- A prompt is a single order ticket handed to an AI. Only what's written on it gets delivered — anything left off never arrives.
- A good ticket spells out what, under which conditions, for whom, and in what shape.
- Leave it vague and the AI fills the gap with whatever's most common. That guess shifts every time, so the answer wobbles too.
- Showing one or two examples of the result you want often beats a long written explanation.
- The ticket gets handed over fresh every time. The entire conversation so far rides along on it.
Contents
1The analogy
Order a rice bowl through a delivery app and the screen asks for more than the dish name. There's a spice level, a portion size, an arrival window, and a box for special requests. Whatever gets typed into that box reaches the kitchen exactly as written. Whatever doesn't get typed never reaches it at all.
Write "make it less spicy" and it comes less spicy. Leave the box empty and it comes however the kitchen normally makes it. The kitchen isn't ignoring a preference — there's no way to know a preference that was never written down.
Write something vague and the kitchen decides on its own. Type "make it however's good" and what arrives is that kitchen's idea of good, which may not match what was actually pictured. A prompt is that same request box.
2In detail
Whatever isn't written never arrives
This isn't the AI missing a hint. Outside of what's written in the prompt, there's no way for it to know the situation at all. Whether the person asking is a student or an office worker, where the writing will be used, how long the answer should run — none of that exists unless it's spelled out.
Ask only "write me a report" and who it's for, how many pages, how formal it should sound are all left blank. The AI fills those blanks with whatever shape it saw most often during training. An answer that feels merely adequate, not quite matched to the actual situation, usually comes from exactly this.
Four things worth putting on the ticket
Spell out what is being asked for with a clear verb. "Summarize this" and "compare these" and "rewrite this" lead to completely different results.
Attach the conditions. Length, tone, what must appear, what has to be left out. A single line like "in three paragraphs," "formal tone," or "explain any jargon" reshapes the result substantially.
Say who it's for. The same explanation shouldn't sound the same aimed at a child as aimed at a specialist.
Specify the shape the answer should come back in. A list or a paragraph, a table or plain prose — naming it means the result can be used as-is.
Vague requests get shaky answers
A request like "make it good" has no standard to aim at. Without one, the AI guesses differently each time, so the same request sent twice comes back different. It's also hard to know what to fix when the answer misses.
Turning a standard into a number or a condition steadies things. "Under five lines" beats "short." "Written for a middle schooler" beats "simple." "Open with one line of apology" beats "polite." There's no need to nail the perfect ticket on the first try — moving each sticking point in the last answer into a written condition, one at a time, is usually faster.
An example moves faster than words
Some tone or format is hard to describe in writing. In those cases, showing one or two finished examples works far better. Attach a sample and say "like this," and the shape carries over without a paragraph of explanation.
Showing what to avoid helps too. But rather than listing only what not to do, pairing it with what to do instead tends to land better.
The ticket gets rewritten every time
Keep a conversation going and it feels like earlier statements stay in effect. What's actually happening is that the whole conversation so far gets bundled onto one ticket and handed over again, fresh, every single time. That's why the earliest parts get pushed out once a conversation runs long.
For a condition that really matters, writing it once at the start and trusting it to stick is riskier than repeating it again at the point where it counts.
3More precisely
A prompt refers to the entire input that goes into a model. It's not just the sentence typed on screen — the system prompt a service inserts ahead of time, the conversation so far, any pasted document, and any retrieved material all merge into a single input. That input is sliced into pieces before it enters the model, and the model produces the pieces that follow it, one at a time, weighing everything placed in front of it as a single block of context.
The analogy breaks down somewhere. A kitchen has a person in it, so an odd request gets a phone call to confirm; a model never asks back, and fills the gap with whatever seems plausible instead. And an order ticket stays on file at the restaurant, but a prompt leaves no trace in the model — it's closer to a single sheet of paper that has to be handed over whole, every time. The conversation still visible on screen is a record kept for the reader's own eyes; on the model's side, it's cargo loaded fresh with every message, with nothing set aside in between.
4Try it yourself
5Common misconceptions
It's easy to think a well-written prompt makes a model do something it couldn't do before, but actually it's drawing out ability the model already has — no new ability gets added.
It's easy to think a longer prompt is always a better one, but actually a short one with clear conditions usually beats a sprawling explanation.
It's easy to think a condition set once holds for the whole conversation, but actually a long conversation pushes the earlier parts out, and that condition disappears along with them.
7One-line summary
In shortA prompt is the order ticket handed to an AI, so only what's written gets delivered, and any box left blank gets filled in on the AI's own judgment.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02