RACE: The technique to enhance prompting

RACE is a simple technique to improve the quality of responses on ChatGPT and other generative AI models. 

RACE is an acronym that stands for:

• Role: the role that generative AI should take.

• Action: what we want this AI to do for us.

• Context: the relevant information for our case.

• Examples: good examples of structures we would like to reproduce.

We can remember this short acronym by thinking of a race towards AI! 

Example of prompting with RACE

So, rather than writing "write a blog post," by applying RACE, we would write instead
"You are a professional writer (Role). Write a blog post to announce our new team (Action). For context, our school, Side School, has just hired 3 key roles for marketing, academia, and IT (Context). Take inspiration from these existing articles: link 1, link 2, link 3 (Examples)."

Importance of examples

Without RACE prompting, generative AI responses are less relevant in all our tests. Providing examples, the fourth step, has the most influence on the quality of the generated responses. 

This is what is called "few-shot prompting" in AI jargon, as opposed to "zero-shot prompting," where no examples are used.

Prompts using RACE

Before Side School, I created an extension called CrowdGPT that aggregates the best prompts from the web. After observing thousands of prompts, I noticed that the best ones had the same structure: this is where the RACE method came from. Here are some examples:

And even, for the smartest ones:

Automatically using RACE

To automatically apply the acronym "Role, Action, Context, Examples," instead of rewriting it each time, I use text expansion on Mac/iPhone. A short tutorial:

On Windows and other OS, you can use this trick in the same way by using a tool like Espanso.

Other techniques complementary to RACE

RACE is a versatile prompting technique that can be supplemented, depending on the case, by the following techniques:

• Prompt Chaining : For complex tasks, instead of using one very long prompt, it is better to break it down into smaller tasks.

• Prompting “divergence then convergence” : For creative tasks, ask the LLM to generate several options, then choose the best one explaining why, before generating a final version.

• Prompting ”Chain-of-Thought” : For tasks with logical reasoning, asking the language model (LLM) to “think in steps” improves final results.

These are techniques we delve into in our Side School trainings.