OpenAI’s Six Steps to Improving Your Prompts to Get Better Results

When using AI tools like ChatGPT, the right prompt will get you the best results. Now, OpenAI has gifted the world a guide on how to improve your prompts.

Quietly published under its website’s documentation section, the prompt engineering guide shares tactics and tips you can use to get better results from large language models like GPT-4.

OpenAI offers six steps, noting that some of the methods can be combined “for greater effect.”

Users can also explore various prompt examples to get the best out of their inputs.

The tactics from the new OpenAI prompt engineering guide can work in tandem with OpenAI’s Cookbook, which contains example code and links to third-party resources on prompting guides and tools as well as video courses.

#1 Write clear instructions

OpenAI said that inputs require users to be specific about their needs. Want short responses? Say so. If you want less simplistic responses, ask for expert-level writing. Using your words will allow a system like ChatGPT to know exactly what you want. After all, these chatbots aren’t quite mind readers, yet.

Credit: AI Business

Here are some tips from OpenAI to improve the clarity of your prompts:

Related:Top 10 Free (or Low Cost) Generative AI Courses

#2 Provide reference text

Systems like OpenAI’s ChatGPT are good – but that doesn't mean they're perfect. Even the most powerful models will sometimes return false responses. OpenAI’s guide said this especially occurs when systems are asked about esoteric topics or for citations and URLs.

The makers of ChatGPT argue that providing reference texts can result in fewer falsities in outputs. To achieve this, OpenAI suggests instructing the model to answer using reference text or with citations from a reference text.

#3 Split complex tasks into simpler subtasks

Instead of uploading a document to ChatGPT and asking it to do everything at once, OpenAI’s prompting guide suggests better results can be achieved by breaking down your desired outputs.

For example, instead of uploading a report and asking the chatbot to summarize the entire thing, break the summaries down into sections and then use those multiple outputs to create a full, more detailed one.

OpenAI also suggests employing intent classification to identify the most relevant instructions.

#4 Give the model time to "think”

OpenAI’s fourth suggestion for improving outputs is simply to have patience. The prompting guide suggests that models make more reasoning errors “when trying to answer right away, rather than taking time to work out an answer.”

Related:Meta Scientist: How Large Language Models Work, AI Summit NY 2023

Ways to improve this that OpenAI suggested include instructing the model to work on its solution to a query before rushing to a conclusion or using a sequence of queries to hide the model's reasoning process. Users could also ask the model to repeat the task, ensuring that it didn’t miss anything on previous passes.

#5 Use external tools

Models like ChatGPT are only as good as the underlying models. However, many systems can be combined with external tools to improve results. Since May 2023, ChatGPT, for example, has had access to a myriad of plugins, with users able to find great restaurants using OpenTable or plan their next vacation with the Kayak extension.

OpenAI’s prompting guide said that plugins or external tools in generate can “compensate for the weaknesses of the model.”

OpenAI’s own Code Interpreter, for example, can improve ChatGPT’s math and code generation abilities.

“If a task can be done more reliably or efficiently by a tool rather than by a language model, offload it to get the best of both,” OpenAI’s guide reads.

OpenAI suggests using embeddings-based search to implement efficient knowledge retrieval or using code execution to perform more accurate calculations.

#6 Test changes systematically

OpenAI suggests users could perform evaluations on the models they’re using to see if the system will give them the desired outputs. The ChatGPT makers suggest users may want to conduct tests like asking their model a collection of different scenarios or questions to test their AI, ensuring it performs well.