What is a prompt? The definition and what really arrives

By Marco Kohns, Co-founder of ENLIX, lecturer in AI and growth

· 12 min read

A prompt is the input you give an AI model to set it a task. The German dictionary Duden lists it as a masculine noun and defines it as an input that serves as a task description for generative artificial intelligence models producing text or images (Duden, "Prompt").

That is half the answer. The other half is in no dictionary entry: your prompt never reaches the model on its own, and what sits inside a prompt that actually produces something usable is almost entirely facts that only you hold.

Both halves are below. What the word means, what a working prompt is made of, counted on a brief from our own course, and what the model reads besides your words before it answers.

What is a prompt?

It is the task description a generative model receives before it produces anything. The model has no other route to what you want. It knows neither your business nor your calendar nor the document on your desk unless that is in the input.

The term is older than ChatGPT. In computing a prompt was for decades the command-line cue, the blinking character with which a program asks you to type something. Since generative models went mainstream, the word in everyday use almost always means the input to such a model.

The load-bearing word in the definition is description. A prompt is not a setting stored somewhere, and not a command that is guaranteed to execute. It is a description, and how complete it is decides how often you have to go back and fix the output.

Where does the word prompt come from?

From English, and before that from French. To prompt someone is to cue them into saying or doing something, which is close to what the input does to a model.

German borrowed the word whole rather than translating it. Duden lists it as a masculine noun, der Prompt, with the genitive des Prompts and the plural die Prompts, gives artificial intelligence as its field of use, and names the English prompt in the sense of Aufforderung as its origin (Duden, "Prompt"). German equivalents exist, among them Aufforderung, Arbeitsanweisung and Eingabe, and none of them caught on.

That matters if you work in a language other than English, because which language you prompt in is a question of its own, not a matter of taste. We wrote a separate piece on using Claude Code in German.

What is a working prompt made of?

Facts only the person commissioning the work holds. We counted the brief that gets typed live in our course, the one that produces a working price calculator page for a wedding photographer: 196 words, 17 pieces of information, not one line of code.

The 17 fall into four kinds. Six are pricing rules from the business's own price list, six are limits, three describe what a good result looks like, and two name the technology already in place.

What is inside the brief that produces a live page in our course

Pricing rules from the business's own list6 itemsLimits on what must not happen6 itemsStatements about the expected result3 itemsFacts about the existing setup2 items

Counted on the wording of the brief from our live build on 9 October 2026. The brief runs to 196 words and contains no code.

What is inside the brief that produces a live page in our course
Pricing rules from the business's own list6 items
Limits on what must not happen6 items
Statements about the expected result3 items
Facts about the existing setup2 items

Not one of those 17 items was available to the model. The base package at 1,400 euros, the 20 percent surcharge for a Saturday between May and September, and the decision that nobody has to leave an email address before seeing a price are in no training set. They are in the price list and in the owner's head.

Which gives the most useful rule of thumb for what belongs in a prompt: everything somebody would need to know to do the job for you without being able to ask you a question. If you want to put such a description together in a structure, use the Prompt Generator. It calls no model. It arranges your own statements into the four blocks goal, context, constraints and example.

Is a prompt the same thing as a search query?

No, and that confusion explains most disappointing answers. A search query looks for something that already exists. A prompt commissions something that does not exist yet.

Search queryPrompt
What happensExisting material is foundNew material is produced
ResultA list of places to lookA finished text, image or file
What works wellA few precise keywordsComplete facts about your case
Too little context gives youFewer hitsA plausible-sounding invention
What you can checkThe source behind the hitOnly the result itself
Why keywords work in search and fail in a prompt.Source: Own comparison, checked 9 October 2026

The last row is the uncomfortable one. In a search you can see where a statement came from. With a produced result you see the result first, and whether it holds up is yours to establish. That is why a prompt should carry, as often as you can manage it, a statement of what a good result looks like. It gives you something to check against.

How does a weak prompt differ from a complete one?

By how much the model has to guess. The weak prompt describes a topic. The complete one describes a result.

The same task, asked twice:

Write me a rejection letter to a candidate.

What comes back is a letter that fits every candidate and none. The model does not know who is writing, what the person applied for, how far they got, or how direct you want to be. So it picks the average of everything it has ever read.

Write a rejection to a candidate who reached our second interview and finished just behind an internal applicant. We are a tax practice with eleven people and the role was payroll administration. 150 words at most, no hedging, the reason is stated rather than talked around, and it closes with an invitation to come back in six months because we expect to be hiring again by then. No phrases along the lines of "we regret to inform you".

The second prompt is not more polite, it carries more information: where the candidate stood, the size and type of the business, the role, a length, two prohibitions and the ending you expect. None of that was available to the model, and every piece of it changes the letter.

That is the whole mechanism. A prompt does not improve because you phrase it more nicely or assign the model a persona. It improves because you close the gaps it would otherwise fill with probabilities.

What types of prompt are there?

Four that come up day to day, and they differ in how far the same words reach.

KindWho writes itWhat happens next
Chat promptyou, in the input boxthe model answers with text
Image promptyou, in an image modelthe model produces an image
System promptthe provider or a developersits in front of each of your messages and governs the whole conversation
Brief for an agentyou, often as a filethe model reads files, writes files and runs commands
Four kinds of prompt, sorted by what happens after you send it.Source: Anthropic, Context windows, Claude Code, How Claude remembers your project

You rarely write the system prompt yourself, but you are bounded by it. It is the reason the same model answers differently inside two products.

The fourth kind carries the most risk, because an agent does not just write, it acts. What an agent is and what it is allowed to do is covered in our piece on AI agents.

What does the model read besides your prompt?

Considerably more than your message, and the providers document it. Anthropic states it plainly for its own API: everything in the request counts toward the context window, namely the system prompt, every message in the conversation including tool results, images and documents, and your tool definitions (Anthropic, Context windows).

Every turn, according to the same page, begins by passing the entire previous conversation history plus your current message back in as input. The model's answer becomes part of the input for the next turn. In a longer conversation your typed question is therefore the smallest part of what is being read.

Two consequences that pay off immediately:

  1. A prompt is never quite on its own. What you ask in turn three is answered against the background of turns one and two. When a conversation has gone the wrong way, a new conversation is often faster than another correction.
  2. More context is not automatically better. On the same page Anthropic names the effect that accuracy and recall degrade as the token count grows, and calls it context rot. Curating what is in context is therefore as important as how much room there is.

A prompt is also not only text. A single request at Anthropic can include up to 600 images or PDF pages, and up to 100 on models with a smaller context window (same source). Whatever you attach is part of the prompt.

Why is the same sentence not the same thing in a chat and in an agent?

Because what happens between your input and your next look changes. In a chat window the model reads and answers. In an agent it reads your project, writes files and runs commands before you see anything.

The same brief, two routes

step comes from you

In a chat window

  1. you type the task
  2. the model reads the system prompt, the history and your message
  3. it writes an answer
  4. you read it and decide what to do with it

2 of 4 steps come from you

As a brief for an agent

  1. you write goal, context, constraints and example
  2. the agent reads the project rules
  3. it reads the files it needs
  4. it writes files and runs commands
  5. you check the result

2 of 5 steps come from you

The number of steps that are yours stays the same. What happens in between without you grows.

Hence the practical difference. In a chat you can start sloppily and sharpen up on the second turn, because nothing happens between turns. With an agent the limits are due in advance, because the steps in between have already run by the time you look again.

Turning a brief like that into something that actually runs is the subject of our piece on vibe coding.

When does a prompt become a file?

As soon as you type the same statements a second time. A chat forgets between sessions, so everything that holds permanently moves out of the input box and into a file that gets read automatically.

Claude Code does this explicitly. Every session begins with a fresh context window according to the documentation, and the instruction files are loaded at the start of every conversation (Claude Code, How Claude remembers your project). The same page records that Claude treats these files as context rather than as enforced configuration: blocking an action reliably needs a hook, not a sentence.

That is the point where typing turns into operations. This article is itself an example. It is produced three times a week from a stored brief, and the rules it is written to are not in that brief but in versioned files beside it. How we cut and maintain such files is covered in our piece on Claude Code skills.

What does a prompt not do?

It does not check itself. A prompt describes what should come out, and it shifts how likely that is. It does not establish whether it happened.

Three limits that no wording moves:

  • A prompt does not grant access. What the model is not allowed to read, it will not read because you asked politely. Missing access is a permissions problem, not a phrasing problem.
  • A prompt does not make a statement true. Produced text can carry sources and figures nobody can verify, because they do not exist. The checking happens outside the prompt. This is the risk the European Commission uses as its example of what staff should know under the AI literacy obligation.
  • A prompt is not a sign-off. Behind the brief for this article sit a test suite and a build that abort the run before anything is published. Asking for the check inside the brief is not enough. It has to stand beside the brief and be able to block.

What you genuinely can improve about the wording itself, and in which order, is covered in our piece on prompt engineering. And if you would rather practise all of this on a project of your own than on examples, that is what the Claude Code System is.

Frequently asked

What is a prompt in simple terms?

The input you give an AI model to set it a task. The German dictionary Duden defines it as an input that serves as a task description for generative artificial intelligence models producing text or images. In practice it is everything you type into the box, plus everything you attach.

Where does the word prompt come from?

From English, where to prompt means to call on someone to say or do something, and before that from French. In computing a prompt was long the command-line cue asking you to type. Duden lists the German noun as masculine, der Prompt, and gives artificial intelligence as its field of use.

What is the difference between a prompt and a search query?

A search query looks for something that already exists and returns places to find it. A prompt commissions something that does not exist yet and returns a result. That is why a few precise keywords pay off in a search, while a prompt pays off for the facts only you hold: your numbers, your limits, the result you expect.

What types of prompt are there?

Four that come up day to day: the chat prompt in the input box, the image prompt for image models, the system prompt written by the provider or a developer that sits in front of every message you send, and the brief handed to an agent, which then reads files and runs commands on its own. The same words reach different distances in the four cases.

What does an AI model read besides my prompt?

At Anthropic everything in the request counts toward the context window: the system prompt, every message in the conversation including tool results, images and documents, and the tool definitions. Your typed message is usually the smallest part of that, and the history grows with every turn.

Written by

Marco Kohns

Co-founder of ENLIX, lecturer in AI and growth

Marco worked as a growth product manager at a Silicon Valley scale-up and has been teaching that way of working ever since. Today he runs ENLIX with Tobias and builds two products of his own on the same systems, which is what the courses open up.

  • Growth product manager at a Silicon Valley scale-up, Series A to B, backed by a16z, General Catalyst and Sapphire, with users in over 100 countries and more than 20,000 cities
  • Peer-reviewed research in the Journal of Business Research on generative AI in growth, with Prof. René Bohnsack. The research began in summer 2022, months before ChatGPT was public
  • Executive education lecturer at Católica-Lisbon, over 10 seminars, more than 1,500 people taught

Keep reading