Large language models are powerful, but they are not mind readers. The same model can give a vague, generic answer or a precise, useful one depending on how you ask. Prompt engineering is the craft of asking well, and it is mostly clear writing plus a bit of experimentation.
Key takeaways
- A prompt is instructions plus context plus format.
- Be specific about who the answer is for, what you want and how it should look.
- Show examples and ask for reasoning steps when accuracy matters.
- Iterate: treat prompts like drafts, and test them.
The anatomy of a good prompt
| Part | Purpose | Example |
|---|---|---|
| Role | Sets perspective and expertise | "You are a senior accountant who writes for freelancers." |
| Task | The specific action | "Explain how to invoice an international client." |
| Context | Background facts the model cannot know | "The client is in Germany; I am in Bangladesh; payment via Wise." |
| Format | Shape of the output | "Give a numbered checklist and a short sample email." |
| Constraints | Limits and rules | "Under 300 words, plain English, no legal advice." |
| Examples | Show the style you want | "Match the tone of the sample below." |
Eight techniques that work
1. Be specific
Vague: "Write about marketing."
Specific: "Write a 150-word LinkedIn post for owners of small restaurants explaining why Google reviews matter, with one practical tip and a friendly tone."
2. Give the model a role
"Act as an experienced UX researcher" changes vocabulary and priorities. Use roles that fit the task.
3. Provide context and source material
Paste the relevant text, data or policy and say "use only the information below". This reduces made-up details. For sensitive content, follow your company's data rules.
4. Show examples (few-shot prompting)
Give two or three input and output pairs, then your real input:
Turn each feature into a benefit.
Feature: 10-hour battery -> Benefit: Work all day without a charger.
Feature: Waterproof case -> Benefit: Use it by the pool without worry.
Feature: One-tap pairing -> Benefit:5. Ask for step-by-step reasoning
For maths, planning and analysis, ask the model to work through the problem and show its steps before giving the final answer: "Think through this step by step, then give your recommendation." It often improves accuracy and makes errors easier to spot.
6. Specify the output format
Ask for tables, JSON, bullet points, headings or word counts. Structured outputs are easier to use in documents and software: "Return JSON with keys title, summary and tags."
7. Break big tasks into steps
Instead of "write a business plan", ask for an outline, then each section, then a review. Chaining smaller prompts gives you control at every stage.
8. Ask the model to critique and improve
"List three weaknesses in this draft, then rewrite it to fix them." Self-review catches problems quickly.
Reusable prompt templates
Summarise a document
You are an analyst. Summarise the text below for a busy manager.
Give: (1) a 3-sentence summary, (2) 5 key points, (3) risks or open questions.
Use only the text provided. Text: """..."""Write a customer email
Write a calm, empathetic reply to this customer complaint.
Acknowledge the problem, apologise once, offer a specific fix and a timeline.
Under 150 words. Complaint: """..."""Plan a content calendar
You are a content strategist for [business] targeting [audience].
Create a 4-week calendar with 2 posts per week.
For each: title, goal, key points, call to action. Output as a table.Debug code
Here is a function, the expected behaviour and the error I get.
List the 3 most likely causes, how to test each and a suggested fix.
Code: ... Error: ...More ready-made prompts are in how to use ChatGPT for business.
Common mistakes
- Too vague, so the model guesses.
- Too many goals in one prompt.
- Hidden assumptions: things you know but did not say.
- No format instruction, so you get a wall of text.
- Trusting the first answer. Check facts and ask follow-ups.
- Pasting confidential data into unapproved tools.
How to test and improve prompts
- Write the prompt and run it on five different realistic inputs.
- Note where outputs fail: wrong tone, missing facts, wrong format.
- Adjust one thing at a time and re-run.
- Save the best version in a shared prompt library with an example input and output.
- Re-test after the model or tool changes.
When building AI features into products, teams add automated evaluations: a set of test inputs with expected qualities, scored every time the prompt changes.
Limits you should know
- Models can state false things confidently. Verify important facts.
- They do not know your private data unless you provide it.
- Long conversations can drift; restart with a short summary.
- Prompts do not replace human judgement, especially for legal, medical and financial decisions.
Take it further
Compare tools in ChatGPT vs Claude vs Gemini, find tools for your work in best AI tools for small business, or turn your best prompts into an automated workflow: see workflow automation ROI and our AI automation service.
