It is the word everyone uses and few people work on. A prompt is simply what you write to the model. Because the technical barrier is zero, people assume there is nothing to learn, and that is exactly where the gap opens between those who find AI disappointing and those who get real value from it.
It is not about magic formulas. It is about precision: the model does not guess what you leave out, and it knows nothing about your situation beyond what you give it.
Definition
A prompt is the instruction you pass to an AI model to obtain a result. It is written in ordinary language and constitutes, together with the conversation history, the entirety of what the model takes into account when answering.
The most useful way to remember it: a model knows about your request only what the prompt contains. Not your trade, not your constraints, not your client, not what you have in mind and did not write down.
A good test before sending: could a competent person who knows neither you nor your company do the job with this information alone? If not, the prompt is incomplete, and the result will be too.
The four parts of a prompt that works
Context
Who is speaking, for whom, in what situation. "I sell an online course to freelancers who are just starting out" changes everything that follows. Without it, the model produces generic output, because that is the average of what it has seen.
Task
A clear verb and a precise object. "Draft", "classify", "compare", "rewrite". One task at a time: a prompt asking three different things gets all three done adequately.
Constraints
Length, tone, format, what to avoid. This is the part most often left out, and the one that makes the biggest difference between a draft and something usable.
An example
A sample of what you expect beats three paragraphs of explanation. It is the most powerful and least used lever: showing a text you liked produces a far closer imitation than describing what you want.
Before and after
| Weak prompt | Worked prompt |
|---|---|
| "Write a LinkedIn post about my product" | "I sell invoicing software to freelancers. Write a 150-word LinkedIn post that starts from a concrete invoicing problem, no emoji, no closing question. Here is a post that performed well, keep this tone: [example]" |
| "Summarise this document" | "Summarise this document in five points, most important first, for someone who has two minutes to decide whether to read the whole thing" |
In both cases the model is identical. Only the instruction changes, and the results are nothing alike.
Best practices
Say what you want, not what you do not want. "Write it short" works better than "do not make it too long". Negative instructions are followed less reliably.
Iterate instead of anticipating everything. Send a simple version, see what is off, fix that specific point. It is faster than writing a perfect prompt first time.
Keep what works. A prompt that produces a good result is an asset. File it somewhere, you will reuse it fifty times. The ones that come back every day belong in a System prompt.
Separate instruction from data. A clean instruction, then the document, with a visible break between the two. The model follows far better when it can see where the instruction ends and the material begins.
Be careful with content you paste without reading. A document from outside can contain instructions aimed at the model, which will override yours. This is known as a Prompt injection: always treat imported content as data, never as an instruction.
Frequently asked questions
Does a longer prompt give a better result?
Only up to a point. Adding useful context clearly improves things, adding filler degrades them: the instruction dilutes inside the Context window. Aim for density, not length.
Should you give the model a role?
The famous "you are an expert in..." has far less effect than claimed on recent models, which do not need an assigned identity to answer well. What genuinely works is stating the audience and the expected level: "for someone new to the subject" or "for a reader who already knows the basics" changes the result far more than an asserted role.
Should you be polite to a model?
Politeness has no measurable effect on quality, but it has an effect on you: framing a request as you would to a competent person naturally pushes you to supply context and constraints. That is where the gain comes from, not from "please".
Does the same prompt give the same result across models?
No, and that is normal: each model has its own writing habits. A well-built prompt still beats a vague one everywhere, so the effort of framing transfers even when the exact output does not.
How do you improve without spending hours on it?
By revisiting your own failed requests rather than collecting prompt templates found online. An instruction that failed for you contains precisely the missing information to add. Our Claude Cowork course works on that method with real cases, including the back and forth that comes with it.