Shape the answer: format, examples and constraints
Get the length, structure and style you want on the first try. Show one good example, say what to do instead of what to avoid, and separate your instructions from your material.
You now know how to say what you want and give the context behind it. The next step is to control what the answer looks like: its length, its structure, its tone and its boundaries. This is where prompts go from "good enough" to "exactly right".
Describe the output you want
If you don't specify a format, the model chooses one, often a long answer with headings and bullet points. Be explicit:
- Length: "in under 100 words", "three short paragraphs", "one sentence".
- Structure: "a table with three columns: option, cost, risk", "a numbered checklist".
- Tone: "warm but direct", "neutral and factual", "like a patient teacher".
- Audience level: "assume the reader has never coded".
Compare renting and buying a flat in Lyon for a couple in
their thirties. Answer with a table (columns: factor,
renting, buying), followed by a two-sentence recommendation.
Say what to do, not only what to avoid
Instructions phrased as "don't" are weaker than instructions phrased as "do". The model knows what to avoid but still has to guess what you want instead.
| Weak | Better |
|---|---|
| Don't be too technical. | Explain it as you would to a smart friend with no science background. |
| Don't make it long. | Keep it under 150 words. |
| Don't use markdown. | Write in plain prose paragraphs. |
A related trick: the style of your prompt influences the style of the answer. A prompt written in friendly prose tends to get friendly prose back; a prompt full of bullet points tends to get bullet points.
Show, don't just tell: use examples
Sometimes the fastest way to explain a style is to show it. This is called few-shot prompting: you include one or more examples of the output you want.
Write a product description for our new water bottle in the
same style as this example:
Example — Travel mug:
"Coffee that's still hot at 3 pm. Fits every car cup holder,
survives every drop, and never leaks in your bag."
Now write one for: a 750 ml insulated steel water bottle.
A few guidelines:
- Start with one example. Add a second or third only if the first doesn't get you there.
- Make examples varied if you use several, so the model copies the pattern, not the specific words.
- Check your examples carefully. Models pay close attention to them and will copy their mistakes too.
Separate instructions from material
When your prompt contains material to work on, like an article, an email or data, make it obvious where the instructions end and the material starts. Otherwise the model may confuse the two.
Clear labels are enough:
Summarise the customer email below in two sentences, then
suggest a polite reply.
Customer email:
"""
Hi, I ordered a blue jacket on 2 May and received a green
one. I need it for a wedding on Saturday…
"""
For long or complex prompts, some people use tags like <email> and </email> around the material. With modern models, clear headings and spacing usually work just as well; use tags when you mix several different pieces of material.
Allow "I don't know"
Models are trained to be helpful, so they tend to answer even when they shouldn't. Give explicit permission to be uncertain:
Based only on the sales data below, what explains the drop
in March? If the data isn't enough to tell, say so and list
what extra information would help.
This one sentence reduces invented explanations more than any amount of "be accurate".
A checklist for this level
Before you send a prompt that matters, check:
- Did I say how long the answer should be?
- Did I describe the structure (paragraphs, table, list)?
- Did I phrase my constraints as "do" rather than "don't"?
- Would an example make the style clearer?
- Is my material clearly separated from my instructions?
- Did I allow the model to say "I don't know"?
Try it now
Write a prompt that produces a table, then a second version that also includes one example of the style you want. Compare the Output format score in the coach.
Sources
- Anthropic, Best practices for prompt engineering (2026)
- PromptBuilder, Prompt engineering best practices 2026