Prompt Engineering
Updated
Prompt Engineering Cheat Sheet
The patterns, structures and fixes that reliably improve output
A one-page reference for writing prompts that work: the anatomy of a strong prompt, the six patterns worth memorising, ready-to-copy templates, and the mistakes that quietly ruin results.
Overview
A prompt is an instruction, not a wish. The single biggest improvement most people can make is to state the ROLE, the TASK, the FORMAT and the CONSTRAINTS explicitly instead of hoping the model infers them.
Anatomy of a prompt
Include these in roughly this order. You rarely need all six, but naming them stops you forgetting the one that matters.
| Part | What it does | Example |
|---|---|---|
| Role | Sets voice, expertise and assumed audience | You are a senior technical editor. |
| Task | The single thing you want done | Rewrite the paragraph below for clarity. |
| Context | Facts the model cannot guess | The audience is non-technical hospital staff. |
| Format | Exact output shape | Return a markdown table with 3 columns. |
| Constraints | Hard limits and exclusions | Under 120 words. No marketing language. |
| Examples | One or two worked cases | Input: ... Output: ... |
Patterns worth memorising
| Pattern | When to use it | Cue phrase |
|---|---|---|
| Zero-shot | Simple, well-known tasks | Just ask directly |
| Few-shot | Output must match a specific style | Here are two examples: ... |
| Chain-of-thought | Multi-step reasoning or maths | Work through this step by step |
| Role prompting | Domain expertise or tone matters | You are a ... |
| Structured output | The result feeds another system | Return valid JSON matching this schema |
| Self-critique | Accuracy matters more than speed | Review your answer and list any errors |
Common mistakes
Each of these silently degrades output. Check them before blaming the model.
- Asking two things at once Split into separate prompts — quality drops sharply on compound asks.
- Describing what you do not want State the positive instruction; negations are followed unreliably.
- No output format If you did not specify the shape, you cannot complain about the shape.
- Burying the instruction Put the task at the top; long preambles push it out of attention.
- Vague quality words "Good", "professional" and "engaging" mean nothing — give a measurable rule.
- No examples for style work One example beats three paragraphs of description.
Prompt templates
Universal rewrite prompt
Editing, summarising, tone changesA dependable starting point for any "make this better" task.
System prompt
You are a meticulous editor. You preserve the author's meaning and voice. You never add facts that are not present in the source.
User prompt
Rewrite the text below for {{goal}}.
Audience: {{audience}}
Tone: {{tone}}
Hard limit: {{limit}}
Return only the rewritten text, with no preamble.
---
{{text}}
Variables
{{goal}}
What you want improved
— e.g. clarity and brevity
{{audience}}
Who will read it
— e.g. first-time customers
{{tone}}
Desired register
— e.g. plain, warm, non-salesy
{{limit}}
A measurable constraint
— e.g. under 100 words
{{text}}
The source text
— e.g. Paste the paragraph here
Example output
Your booking is confirmed for Tuesday at 2pm. We will send a reminder the day before. If you need to change it, reply to this message and we will sort it out.
If the result drifts, add: "Do not change any names, dates or numbers."
Structured extraction prompt
Pipelines, data entry, form fillingTurns messy text into machine-readable JSON.
System prompt
You extract data. You return valid JSON only, with no commentary and no markdown fences.
User prompt
Extract the fields below from the text.
If a field is absent, use null. Never guess.
Schema:
{{schema}}
Text:
{{text}}
Variables
{{schema}}
The JSON shape you want
— e.g. {"name": string, "date": string|null}
{{text}}
Source content
— e.g. Paste the email here
Example output
{"name": "Priya Raman", "date": "2026-03-14"}
Pair this with the API's JSON mode where available — the instruction alone is not a guarantee.
Frequently asked questions
Does prompt length matter?
Only insofar as it costs tokens and can push your instruction out of focus. A precise 60-word prompt usually beats a rambling 400-word one.
Should I use "please" and "thank you"?
It makes no measurable difference to quality. Use whatever register you are comfortable with.
Why does the same prompt give different answers?
Most models sample randomly. Set temperature to 0 (or the lowest available) when you need repeatable output.
Resources
Was this cheat sheet useful?
Comments
No comments yet — be the first.
Keep going
Browse all
More cheat sheets
Need a different cheat sheet?
Tell us what you would like to see and we will build it — free.
Request a cheat sheet
