OpenAI has just published the authoritative guide to doing better prompt engineering with ChatGPT.
Called GPT best practices, the article outlines six strategies to get better results with ChatGPT using GPT-4. Each strategy is accompanied by a number of different tactics to operationalize each strategy.
The GPT4 strategies are as follows:
Write clear instructions.
Tactics to write clearer include: using more detail to get more relevant answers, asking the model to adopt a persona, and specifying the steps you want GPT-4 to take to do what you want.
Provide reference text.
To cut down on hallucinations, it helps to instruct GPT-4 to use a reference text when answering. Says OpenAI: “In the same way that a sheet of notes can help a student do better on a test, providing reference text to GPTs can help in answering with fewer fabrications.”
Split complex tasks into simpler, smaller tasks.
You can do this in a number of ways, including: prompting GPT-4 to use intent classification to identify the most relevant instruction, summarizing or filtering previous conversations with GPT-4, and summarizing long documents in smaller pieces.
Give it time to “think.”
This means telling GPT-4 to give you its chain of reasoning before computing an answer. This gives the system the ability to reason more reliably.
Use external tools.
OpenAI recommends you “compensate for the weaknesses of GPTs by feeding them the outputs of other tools.” Examples include using a text retrieval system to help GPT-4 learn about documents or a code execution engine to help it do math and code.
Test changes systematically
OpenAI warns “In some cases a modification to a prompt will achieve better performance on a few isolated examples but lead to worse overall performance on a more representative set of examples.” So you need to test regularly to improve performance.
“These may be worth trying with other tools, too,” says Marketing AI Institute founder/CEO Paul Roetzer on Episode 50 of the Marketing AI Show. While OpenAI specifies all of these are designed for GPT-4, the logic may also benefit you as you use tools like Bard and Claude.
It’s imperative to try prompting multiple times, says Roetzer. If you get a bad output on your first prompt, it may not be GPT-4’s fault—you may just need to take a different approach or provide more detail.
Last, but not least, think of AI prompting as a critical skill you need to move forward in your business—because it is.
Large language models are going to be a critical part of every marketer’s job, every business.