← All posts
AI FluencyTools

The four ingredients of a good prompt

By Zoe Wong

If you’ve looked up "how to write a good prompt" online, you’ve probably found 20-step frameworks, exotic role-play tricks, and prompt libraries that read like Excel spreadsheets. Most of it is noise.

Strip it down and a good prompt has four ingredients.

1. Role

Tell the model who it is in this conversation. Not because the model "becomes" anything — it doesn’t — but because giving it a role narrows its working vocabulary in useful ways.

"You are a senior HR business partner reviewing draft performance reviews."

2. Task

Be explicit about what you want done. Verbs first. One concrete task per prompt.

"Rewrite the review below to be clearer, more specific, and structured around three competencies: ownership, communication, and delivery."

3. Context

Give the model what it can’t guess: the audience, the constraint, the situation. A surprising amount of bad output is just the model filling in context you forgot to provide.

"This is for a senior engineer with 8 years of experience. Performance is strong overall but communication during incidents is the area for growth."

4. Format

Tell the model how the output should be shaped. Markdown headings? Three bullets? A table? Specifying this prevents the "wall of text" failure mode.

"Reply with: (1) a one-sentence summary, (2) three competency bullets with specific evidence, (3) one clear development action."

Put it together

Role — You are a senior HR business partner reviewing draft performance reviews.
Task — Rewrite the review below to be clearer, more specific, and structured around three competencies: ownership, communication, and delivery.
Context — This is for a senior engineer with 8 years of experience. Performance is strong overall but communication during incidents is the area for growth.
Format — Reply with: (1) a one-sentence summary, (2) three competency bullets with specific evidence, (3) one clear development action.

[Draft review]
...

That’s it. Role, task, context, format. Everything else — chain-of-thought, few-shot examples, exotic system messages — is an optimisation you only need once you’ve mastered these four.