How to Write Better Prompts: A Practical Guide

Most weak prompts fail the same way: they describe a topic but not a task. The model is left to guess who the output is for, how long it should be, what it must include and what it must avoid — and guesses generically.

The six elements

Context. What situation is this output for? A cover letter prompt that mentions the job, the company and your background produces something usable. One that doesn’t produces a template.

Role. “You are an experienced UK employment solicitor” changes the vocabulary, the caution level and the structure of the answer. Roles work because they compress a whole set of expectations into one line.

Structure. If you want sections, say so. If you want a numbered process, say so. Models follow explicit structure far more reliably than implied structure.

Constraints. Word counts, banned phrases, required inclusions. Constraints do more work than instructions because they are checkable.

Output format. Table, JSON, markdown, plain paragraphs — name it. Unspecified format is the most common reason output needs redoing.

Reasoning instructions. For anything analytical, telling the model to diagnose before answering (“state your assumptions first”, “identify the root cause before proposing a fix”) measurably improves accuracy.

Before and after

Before: “Write an email to my landlord about the broken boiler.”

After: “Write a firm but polite email to my landlord. Context: the boiler has been broken for 9 days, I reported it on 3 July, and I have two young children. Reference the landlord’s repair obligations in England. Ask for a repair date within 48 hours. Under 200 words, plain English, no legal threats.”

The second version isn’t longer because detail is decoration — every added line removes a guess the model would otherwise make.

Test it on your own prompt

The fastest way to internalise this is to score a prompt you actually use and see which of the six elements it’s missing.