It seems like every other technology PR person I meet in 2024 is an expert in prompt engineering, the process of writing an instruction that can be interpreted and understood by a generative AI model. But how many of those people can describe their knowledge and expertise in a way that can be easily remembered and applied?

BY RICHARD STONE

People who aren’t very interested in physics often claim that Einstein once said, ‘if you can’t explain it simply you don’t understand it well enough’. While that might not fully hold water when we are talking spin-statistics theorem, when it comes to writing a prompt for ChatGPT, it’s definitely true.

Fortunately, the good people at StructuredPrompt.com have come up with an excellent way of both describing a well written prompt and remembering exactly what it should contain.

SAY HELLO TO TRACI

Meet TRACI, their neat little acronym that means you won’t ever need to type ‘answer all my emails for me’ into a generative AI ever again:

• Task – The general activity ChatGPT is being asked to perform
• Role – The perspective or persona ChatGPT should adopt in generating the response
• Audience – Who the response is intended for
• Create – The format or medium of the requested output
• Intent – The underlying purpose or goal of the generated text

In addition to including these five basics, it’s worth noting the way that Structured Prompt recommends that you, well… structure your prompt. They argue that each of the subsections of TRACI should be provided with rules, allowing the GPT to understand exactly what you are asking of it. For instance, you might write:

Task: Write a 450-600 blog post about Prompt Engineering’s TRACI acronym, highlighting how to use it and pointing out a couple of interesting ways that it works.

Task_rules:

• Avoid technical or marketing jargon
• Include reasons why each interesting facet of TRACI benefits the reader
• Use a practical and worked example
• Don’t use that cliched hack of making the worked example the actual piece you are writing. You might think you are being funny and meta when you do that, but everyone else does it as well.

For the record, I didn’t get ChatGPT to write this article and I won’t be including a version of this piece, written by Chat GPT, as an addendum so you can see the difference between the two. Why? Because its not 2022, that’s why.

ENGINEERING PROMPTS

One of the interesting points here is that Prompt Engineering chooses to write the rules as if they were fake code, by including underscores between the words. They also put their instructions in bullet point format.

At first, I genuinely thought they were doing this to look cool. Given some further thought though, the underscore is itself a token that makes up part of the prompt you are submitting to the AI. So, if you use an underscore to provide clarity in your writing, you are likely to get a clearer prompt as a result.

Similarly, the bullet point itself, which is to say the little dot at the start of your sentence, represents a distinct token. In fact, in some cases, the space between the little dot in the bullet point and the first letter of your sentence is an entirely separate token itself. And, if you submit two different strings of tokens to a generative AI, you will get two different sets of results.

I’m uncertain, and don’t have enough data to prove, whether the Structured Prompt approach works better on a technical level by providing enhanced clarity, but it’s clear that, by inputting a different token string, it will work differently.

ENGINEERING YOURSELF

However, I don’t really think it’s about engineering ChatGPT, or your preferred generative AI at all. Having worked with hundreds of copywriters in my career, and made a million and one copywriting mistakes myself, I think it’s about engineering you – the prompt writer.

The thing that TRACI reminds me of most, is the advice we give to brand new account executives when they join Stone Junction and start work on their first piece of content. We tell them to know what the client wants to achieve, write with the reader in mind, and know what you want the reader to do after they have read the piece.

With that in mind, it’s no wonder that everyone is an expert in prompt engineering. It turns out that a good prompt uses all the same rules as a good piece of copy.