Getting Started with Prompts

A prompt is not a "well-written" sentence.
It is an operational instruction designed to reduce ambiguity.

If the model gets it wrong often, it is not because it is bad.
It is because the prompt is vague.

1. Start Simple... Then Refine

A good prompt answers 3 questions:

• What? (the task)
• Why? (the objective)
• With what constraints?

Examples:

• Bad: "Give me a business idea"
• Good: "Suggest 3 B2C business ideas, gross margin >40%, launchable in 30 days, max budget 2,000 EUR, no physical logistics."

Tip: start broad, then add one constraint at a time. Too many constraints from the start make the response rigid.

2. Give the Model a Clear Role

A vague role produces a vague response.

• Bad: "Act like an expert"
• Good: "Act as a B2B product manager with 10 years of experience, used to early-stage MVPs."

The role frames the vocabulary, assumptions, and level of detail.

3. Specify the Expected Deliverable

Do not let the model guess the format.

Specify:

• output type (list, table, plan, text)
• length
• level of detail

Example: "Answer as a numbered list, 1 sentence per point, no introduction."

4. Use Examples (Few-Shot)

An example is worth more than a rule.

Example: "Here is a response I consider good: [EXAMPLE]. Do the same on a different topic."

Models imitate better than they understand.

5. Always Challenge the Response

A response is never the truth.

Systematically add:

• "What implicit assumptions did you make?"
• "What is the opposing view?"
• "In what cases would this answer be wrong?"

6. Improvement Loop

Simple process:

1. Raw prompt
2. Response
3. Clarification
4. Simplification
5. Stable version

The perfect prompt does not exist. A good-enough, reusable prompt does.