Prompt Engineering: How to Write Better AI Prompts

What is it?

A prompt is the input you give a model. Prompt engineering is the craft of designing that input so the output is useful and dependable. It is less about secret phrases and more about the same things that make instructions to a human colleague work: clarity, context and a definition of "done".

A model cannot read your mind or see your files. It only has the text (and any other inputs) in the request, so anything not stated is guessed.

Why does it matter?

Vague prompts produce generic or off-target answers; precise prompts produce answers you can use with little editing. For developers, prompt quality directly affects the reliability of an application: a prompt is part of your program, and small wording changes can change behaviour. Treat prompts as something to version, test and review, not as throwaway text.

How does it work?

A dependable structure covers these elements (use the ones that apply):

1. Task — what to do, in one clear sentence.
2. Context — who it is for, background, the relevant source text. Put long source material clearly delimited (for example between labelled markers).
3. Constraints — length, tone, what to avoid, what to do if information is missing.
4. Output format — bullets, table, JSON with named fields, headings.
5. Examples — one to a few input→output pairs ("few-shot") when the format or style is hard to describe. Zero-shot means no examples.
6. A way out — permit "I don't know" or "not in the provided text" so the model is not pushed to invent an answer.

Additional techniques:

• Ask for reasoning on hard problems — telling the model to work through steps often helps with multi-step logic or arithmetic, though results vary by model, and many current models reason by default.
• Break big tasks into steps — extract first, then analyse, then write.
• Iterate — change one thing at a time and compare outputs on several test inputs.

Example

Weak prompt:

Summarise this article.

Stronger prompt:

You are helping a busy product manager.
Summarise the article between the <article> tags in exactly 3 bullet points,
each under 20 words. Then list any decisions the article says must be made.
If the article contains no decisions, write "No decisions mentioned."
Use only information in the article.

<article>
...text here...
</article>

The second version defines audience, length, format, a fallback, and a grounding rule.

When should I use it?

Invest in prompt design whenever you will reuse a prompt, when output feeds another system, or when errors are costly. For a one-off casual question, a short natural request is fine; follow up to refine. When prompting alone cannot supply the knowledge or consistency you need, move to What is RAG (Retrieval-Augmented Generation)? (to add information) or What is Fine-Tuning an AI Model? (to change behaviour) — see RAG vs Fine-Tuning: Which Should You Use?.

Common mistakes

See Common Prompting Mistakes (and How to Fix Them) for a fuller list. The usual culprits are vague goals, missing context, no output format, contradictory instructions, and testing on a single example. Remember too that a prompt is not a security boundary — see What is Prompt Injection?.

Related concepts