In late 2023 OpenAI published a prompt-engineering guide built around six strategies for getting better results from its models. The advice holds up well, and it applies just as much to Claude, Gemini and other assistants. This guide explains each strategy with a before-and-after example, then covers what has changed now that “reasoning” models think through problems on their own.
Key takeaways
- Be specific: say who the audience is, what format you want and how long it should be.
- Give the model the source material instead of relying on its memory.
- Split big jobs into steps, and check each step.
- With today’s reasoning models you rarely need to say “think step by step”. Focus on a clear goal, context and constraints instead.
- Test prompt changes on a set of real examples before rolling them out.
1. Write clear instructions
Models can’t read your mind. The more precisely you describe the task, audience, format and constraints, the less the model has to guess.
| Weaker prompt | Stronger prompt |
|---|---|
| Summarise this report. | Summarise this report for a busy CFO in 5 bullet points, each under 20 words. Lead with the financial impact and flag any risks. |
- State the role or perspective (“You are reviewing this as a GST consultant”).
- Specify the output format: bullets, a table, JSON, an email.
- Give length limits and tone.
- Show an example of the output you want when the format matters.
2. Provide reference text
Models can invent facts, especially about niche topics, recent events or your own business. Paste the relevant document and tell the model to answer only from it, and to say so when the answer isn’t there.
| Weaker prompt | Stronger prompt |
|---|---|
| What is our refund policy for damaged items? | Using only the policy text below, answer: what is our refund policy for damaged items? If the policy doesn’t say, reply “Not covered in the policy.” [policy text] |
3. Split complex tasks into simpler subtasks
One giant prompt that asks for research, analysis and a polished report tends to do each part poorly. Break it into steps, and feed each step’s output into the next.
- First, extract the key facts from the source.
- Then, analyse them against your criteria.
- Finally, draft the output from the analysis.
This is also how automation workflows use AI: one model call to classify an email, another to draft a reply, a rule to decide whether a person reviews it.
4. Give the model time to “think”
The original advice was to ask for reasoning before the final answer, for example “work out your own solution first, then compare it with the student’s answer”. With older models this made a real difference on maths and logic problems.
What changed: reasoning models (the “thinking” modes in ChatGPT, Claude and Gemini) already reason internally before answering. For these, OpenAI’s own guidance is to keep prompts direct and not to over-instruct the reasoning. Give a clear goal, context and success criteria, and let the model work.
5. Use external tools
Models are weak at exact calculation, live data and searching large document sets. Give them tools instead: a code interpreter for calculations and data analysis, web search for current facts, or retrieval over your own documents. In automation platforms, those “tools” can also be APIs or RPA bots; see where AI agents fit.
| Weaker prompt | Stronger prompt |
|---|---|
| What was our average order value last quarter? [pastes 5,000 rows] | Use the code tool to calculate average order value by month for Q3 from the attached CSV, and show the code you ran. |
6. Test changes systematically
A prompt that works on one example can fail on the next. If a prompt is part of a business process, keep a small test set of 20–50 real inputs with known good outputs, and re-run it whenever you change the prompt or the model.
- Track accuracy, format errors and edge cases.
- Change one thing at a time.
- Keep the prompt that wins on the test set, not the one that looked best once.
A reusable prompt template
| Part | What to include | Example |
|---|---|---|
| Goal | What you want and why | Draft a reply to this customer complaint |
| Context | Background and source material | Order details, our returns policy, the customer’s message |
| Constraints | Tone, length, rules, what to avoid | Under 120 words, apologetic, no discount offers |
| Format | Exact output structure | Subject line, then the email body |
| Check | How to handle uncertainty | If information is missing, list the questions to ask |
For more on safe use, see our dos and don’ts of generative AI, and for ideas on where to apply these techniques, 100 ChatGPT use cases.
Common prompting mistakes to avoid
- Vague success criteria. “Make it better” gives the model nothing to aim at. Say what better means: shorter, more formal, fewer jargon terms.
- Burying the instruction. Put the task at the start or end of a long prompt, not in the middle of pasted material. Wrapping the material in clear markers such as quotes or tags helps.
- Contradictory rules. “Be detailed” and “keep it under 50 words” in the same prompt force the model to guess which matters more.
- Asking for facts it cannot know. Your internal prices, last week’s sales or a client’s history must be supplied, not recalled.
- Accepting the first answer. Ask follow-up questions, ask the model to critique its own draft, or request three alternatives and pick one.
Applying the strategies in automation workflows
When a prompt runs inside an automated process rather than a chat window, the strategies matter even more, because no one reads every output. A typical email-triage flow in Make, n8n or Power Automate shows how they fit together:
| Step | Strategy used |
|---|---|
| Classify the email into one of five fixed categories, returning JSON only | Clear instructions, exact format |
| Look up the customer record and pass it in | Reference text |
| Draft a reply in a separate call, using the category and record | Split into subtasks |
| Validate the JSON and route low-confidence cases to a person | Systematic testing and checks |
For a wider view of how these pieces connect, see our workflow automation tools comparison.
Frequently asked questions
Do these strategies work for Claude and Gemini too?
Yes. They are general principles for working with large language models, not OpenAI-specific tricks.
Should I still say “think step by step”?
For reasoning models, usually not; they already reason internally, and extra instructions can make results worse. For quick, non-reasoning models it can still help on multi-step problems.
How long should a prompt be?
As long as it needs to be to remove ambiguity. Detailed context and examples generally beat short, clever phrasing.
Leave a Reply