The Complete OpenAI Prompt Engineering Guide

OpenAI's guide is built around six core strategies, each with specific tactics you can apply immediately: 1) Write clear instructions, 2) Provide reference text, 3) Split complex tasks into subtasks, 4) Give models time to think, 5) Use external tools, 6) Test changes systematically. Master these in order and you'll get dramatically better results from any GPT model.

The number one mistake is being vague. GPT can't read your mind. Include details about the desired context, outcome, length, format, and style. Ask the model to adopt a persona for domain expertise. Use delimiters (triple quotes, XML tags, markdown headers) to separate instructions from content. Specify exact steps for multi-part tasks. Provide examples of desired output. Instead of 'summarize this', say 'summarize in 3 bullet points, each under 20 words, focusing on business impact.'

Write a product description for our AirPods Max competitor.

Constraints:
- 80-100 words
- Target audience: audiophiles aged 25-40
- Highlight: 40-hour battery, spatial audio, titanium build
- Tone: premium but not pretentious
- End with a single-line CTA
- Do not use superlatives like 'best' or 'revolutionary'

When you need factual accuracy, give the model source material and instruct it to answer using only that text. This is the simplest form of RAG (Retrieval-Augmented Generation) and dramatically reduces hallucination. You can instruct the model to answer with citations from the provided text, or to say 'I don't know' if the answer isn't in the source material.

Answer the customer's question using ONLY the policy document below. If the answer isn't in the document, say 'I don't have that information in the current policy document.'

<policy>
{{REFUND_POLICY}}
</policy>

Customer question: What is the refund policy for digital purchases?

Complex tasks have higher error rates than simple ones. Instead of one massive prompt, break the work into a pipeline of focused steps. Use intent classification first to route to the right instructions. For long conversations, summarize previous dialogue to maintain context without hitting token limits. Think of it like software engineering: decompose the problem.

Step 1 - Classify the inquiry:
'Categorize this customer message into: billing, returns, technical, account, or general.'

Step 2 - Route to specialized prompt:
Each category has its own focused prompt with specific instructions, tone, and available actions.

Step 3 - Generate response:
The specialized prompt handles the actual response with full context for that category.

Chain-of-thought prompting is one of the most powerful techniques available. Ask the model to work through its reasoning before giving a final answer. This is especially effective for math, logic, code analysis, and multi-step problems. You can use 'inner monologue' to hide the reasoning from end users, or use a sequence of queries where the model's thinking is kept separate from the final output.

First, work out your own solution to the problem. Then compare your solution to the student's solution and evaluate if it's correct. Do not decide if the student is correct until you have solved the problem yourself.

Problem: I need to pay $100 in total. I've paid $40 and $30. How much do I still owe?
Student answer: $20

Don't force the model to do everything. Compensate for weaknesses by feeding outputs of other tools into the model. Use text retrieval (RAG) to inject relevant documents dynamically. Use code execution for accurate math and data processing. Give the model access to APIs through function calling. Embeddings-based search can efficiently find relevant knowledge at runtime.

If a prompt modification improves one example, that doesn't mean it's better overall. It might actually hurt other cases. Define a comprehensive test suite ('eval') that covers the full range of expected inputs. Evaluate outputs against gold-standard answers, either automatically or with a separate model evaluation step. Without systematic testing, you're flying blind.