Seven prompt patterns that actually help

“Prompt engineering” sounds like a doctoral thesis. It is closer to “phrasing things clearly.” Anyone who practices the following seven patterns for two hours gets more out of any AI model, and quickly notices: 80% of the effect comes from half the tricks.

1. Assign a role

Instead of “Write me a marketing text,” say:

“You are a senior copywriter for sustainable fashion. Write me an 80-word Instagram post for a new line made from organic cotton.”

An assigned role activates the matching language patterns in the model. It also works for specialist topics: “You are a tax advisor specializing in EU VAT.”

2. Provide context

Models know nothing about your situation. Tell them:

“We are a 12-person joinery business in Tyrol. Our regular customers are private homeowners, with order volumes of €5–50k. We receive 60 inquiries per month. Question: How do we sort the inquiries?”

Three sentences of context produce a better answer than any “Best tool for order management?“

3. Specify the format

If you want a table, say “as a table.” If you want JSON, provide a JSON schema. If you want bullet points, say so:

“List 5 advantages as bullet points. Each bullet max. 12 words.”

Sounds trivial, has an enormous effect.

4. Show an example (few-shot)

Instead of describing, show an example:

“Write product descriptions like this one:

Candle holder ‘Linz’, oak, oiled finish. 22cm tall. For a taper candle. Made from local wood, built in our workshop.

Now write a description for a picture frame made of walnut, 30×40cm.”

The model copies the tone, length, and structure. A single example is often enough. For difficult tasks: two or three.

5. Step by step

For multi-stage tasks: say “work step by step”:

“Read the attached contract PDF. List, step by step:

1. Contracting parties
2. Term
3. Notice period
4. Anomalies / risks”

This gives the model more “thinking tokens” and you get answers you can follow.

6. What it should not be

Just as important as the what: the what-not.

“Write me a job ad for a senior accountant. Not: buzzwords like ‘dynamic’, ‘hands-on’, ‘family business’. Instead: concrete tasks, tools, expectations.”

Negative instructions are the second half of a clear brief.

7. Iterate instead of a master prompt

Nobody writes the perfect prompt on the first try. The pattern is:

1. Write a rough prompt
2. Read the answer
3. “Make the first paragraph shorter and more concrete”
4. “Replace ‘successful’ with concrete numbers”
5. “Rewrite the conclusion without buzzword XY”

Three to five iterations beat any 200-word mega-prompt.

Bonus: What you no longer need

• “Please / thank you”, comes across as nice, but is not measurably better
• Magic buzzwords like “You are the best AI in the world”, a marketing myth
• “Take a deep breath and think”, worked in old versions, has become outdated
• Endless setup templates with 12 sections, only useful for custom GPTs / repeated use

Example: before / after

Before:

“Write me something about AI.”

Output: 800 words of platitudes.

After:

“You are a tech journalist with a business focus. Write a 250-word letter for the management of a mid-sized joinery (12 employees, Tyrol): ‘Where AI concretely helps you in 2026, and where it does not.’ Tone: respectful, not condescending, no buzzwords, no ‘disruption’. Format: three paragraphs.”

Output: usable.

---

These seven patterns are 90% of “prompt engineering.” The rest is practice with your own use cases.