Examples · General
Prompting 101: how to ask well
The same question asked badly and asked well, side by side. The four parts of a good request (role, context, format and an example) and what each one changes in the answer. No code.
Beginner · 25 min
What you need
Only the chat: no code. We tried every request on this page with Gemma 4 26B, the default model.
The same question, twice
The situation: next week you have a short oral test on cloud computing, and you always mix up IaaS, PaaS and SaaS.
The bad request:
Explain cloud computing.
The good request:
You are a teacher who explains technology to beginners.
I am a first-year student. Next week I have a short oral test on cloud computing, and I always mix up IaaS, PaaS and SaaS.
Explain the difference in a table with three columns: the model, what I manage myself, an everyday example. Under the table, give me one sentence to remember each model. No more than 150 words in all.
One row, to show you what I mean:
| SaaS | nothing: I only use the software | Gmail |
Side by side, part by part:
| Bad | Good | |
|---|---|---|
| Role: who answers | — | a teacher who explains technology to beginners |
| Context: who you are, what it is for | — | a first-year student, an oral test next week, mixes up IaaS, PaaS and SaaS |
| Task | explain cloud computing | explain the difference between the three |
| Format: the shape of the answer | — | a table with three named columns, one sentence per model, 150 words at most |
| Example | — | one row of the table, written the way you want it |
What came back
| Bad | Good | |
|---|---|---|
| Length | about 600 words | about 140 words |
| What was in it | an analogy with electricity, the three models, public, private and hybrid cloud, the benefits, and in one run a table cloud versus your own servers | the table you asked for, three rows, and one sentence per model |
| Ready for the test? | the answer is in there somewhere: you have to dig it out | yes: read it twice and close the chat |
The bad answer was not wrong. It answered a question nobody asked: “tell me everything”. The good one was this, with a line of encouragement before and after it. Running it again gave the same shape with other words: an empty house, a furnished studio and a hotel, in place of the kitchen and the restaurant.
| Model | What I manage myself | Everyday example |
|---|---|---|
| IaaS | Everything: OS, apps, and data | Amazon Web Services (AWS) |
| PaaS | Only my code and my data | Google App Engine |
| SaaS | Nothing: I only use the software | Gmail |
- IaaS: You rent the raw building materials to build whatever you want.
- PaaS: You rent a fully equipped kitchen to focus only on cooking.
- SaaS: You simply order a meal from a restaurant.
What each part does
- Role sets the level and the tone. With “a teacher who explains technology to beginners”, the sentences to remember were everyday images: a house, a kitchen, a restaurant. With “a senior cloud engineer” and the same context, the examples became more technical, Amazon EC2 and Heroku, and in one of our two runs “middleware” appeared in the table and the sentences were about computers and code.
- Context is what the model cannot guess: who you are, what you need it for, what you already know. “I always mix up IaaS, PaaS and SaaS” is why the answer talked about those three and nothing else.
- Format is the shape: a table, a list, how long, which language. Without it, the model picks the shape, and it usually picks long.
- An example shows instead of telling. The SaaS row of the answer is the one we wrote, and the other two rows copied its style.
You do not need all four every time. A quick question needs none. When the answer is long, vague or off target, check which part is missing.
Keep talking
The chat remembers the conversation, so the next request can be short. After the good answer we wrote:
Good. Now the same for public, private and hybrid cloud.
The answer came back in the same shape: the same table, one sentence per model. You do not need to repeat the format, only what changes. Two more words work the same way: “Shorter.” gave the same table in 99 words instead of 145, and “In Italian.” the same table in Italian.
When you do not know what to say
Let the model ask:
I have to give a five-minute talk to my class about my internship. Before you write anything, ask me the questions you need, one at a time, and wait for my answer.
When we tried it, the first “question” was really two at once: the name of the company and your job title. Answer both, or write “one at a time, please”. Do not give names of people, or anything the company would not want in a talk.
Exercise
- Take a question you asked the chat this week. Write it again with a role, the context, the format and an example. Compare the two answers.
- In the good request, change only the context: you are preparing a job interview, not an oral test. What changes?
- Change only the format: five bullet points; or an explanation you can say aloud in thirty seconds.
- Write the good request in Italian: the table comes back in Italian.
- Send the good request to Apertus 70B or Ministral 3 14B. Is the table still a table? When we tried, Apertus broke it in all three runs: a fourth column and 446 words, a table without its header, a header without the line under it that Markdown needs. Ministral kept the table every time, but some of its sentences to remember made no sense. A request that works on one model has to be tried again on another.