The anatomy of a useful prompt
Six ingredients, used as needed. This is a way of thinking about specification, not an incantation you must recite.
What you will be able to do
Diagnose why a prompt underperformed and fix the specific missing ingredient.
Before you start
- Access to any assistant.
The six ingredients
Role: the stance to take. Useful when expertise or register matters, unnecessary otherwise.
Task: the actual job, stated as a verb. The one ingredient that is never optional.
Context: what the model needs and cannot know — your audience, your constraints, the source material.
Constraints: length, tone, what to avoid, what must be included.
Output: the shape you want back. A table, a list, a draft email, JSON.
Evaluation: how you will judge it, stated in advance. The most skipped and most useful.
Short is often correct
"Fix the grammar in this paragraph" needs no role, no constraints and no evaluation criteria. Adding them wastes your time and clutters the context.
Reach for the full structure when the task is ambiguous, the output has a required shape, or you have already been disappointed once.
Diagnosis by ingredient
Wrong register: the role or constraints were missing. Generic content: context was missing. Right idea, wrong shape: output was unspecified. Plausible but wrong: you needed sources, not recall. Nearly right: state the evaluation criteria and ask it to check its own draft against them.
What good looks like
- You can name which ingredient was missing when output disappoints.
- Your short prompts stay short.
- Your evaluation criteria are written before the output arrives.
What it cannot do
- No prompt makes a model reliable on facts it was never given.
- Phrasing that helps one model may not transfer identically to another.
When it goes wrong
- Long prompts that still miss.
- Length is not specification. Check whether the actual task verb is clear.
- Copying prompt templates that do not help.
- Templates carry someone else's context. Keep the structure, replace the content.
Privacy and data handling
Everything in the prompt goes to the provider. Anonymise before pasting, not after.
Sources2 referenced
- Prompt engineering overview
Anthropic · official guidance · link checked 2026-09-06
- Prompt engineering guide
OpenAI · official guidance · link checked 2026-09-06
Links are checked at build time. If one has moved, tell us.
Read next
Iterate, evaluate and improve
The difference between someone good at this and someone frustrated by it is usually a habit of changing one thing at a time.
Checked 2026-09-06
Structured outputs and reusable prompt patterns
When output feeds a spreadsheet, a script or a colleague, its shape stops being cosmetic.
Checked 2026-09-06