AI is weird. In one way or another, we’re all in agreement about that.

Whether it’s because of its mind-blowing powers of recall, it’s vaguely alarming Skynet energy, or because of its strange predilection for just making stuff up. Weird.

But the weirdest part for me is that here’s this system that’s more knowledgeable than any human, can delve through centuries of history, science, politics, economics, any subject you like in seconds… but try getting it to write like a human…

And that’s where I’ve seen a lot of people drop off. Disappointed. They see that AI isn’t going to write what they’re after at the drop of a hat. They’d rather not have “fostering growth”, “harnessing the power” or “paving the way for enhanced performances and transformative breakthroughs” anywhere near their blogs or social posts. And it isn’t going to add playful phrases like “at the drop of a hat” to sentences to make the writing seem more human either. So they give up, underwhelmed, and write it themselves.

But in these early days of working with AI tools, I’ve found there are two Ps that are more important than anything else: perseverance and prompting. It’s these two that will help you discover the real benefits of AI.

You’ll need to bring the first yourself. But on the second, we’ve got a few pointers.

Prompting techniques: an explainer

If you or your business haven’t used AI systems a whole lot yet, ‘prompting’ is simply the messaging you use to guide the AI into producing whatever it is you want it to produce. It’s your way of describing to the AI your desired output, and it can be as simple as a single question or as complex as a 20,000-word set of instructions, incorporating a tone of voice guide and a host of examples (as language nerds, they’re our favourites).

Trying to work out the most accurate way to prompt has inspired everything from social posts to academic papers, with ever weirder and wilder ideas to get the AI to do exactly what you want it to do. It’s just as much art as science, with right-brained thinking like offering the AI bribes or giving it in-depth biographies to get better outputs.

The fun part: you can try different techniques, get creative, mix and match. Whatever achieves the best outputs.

Ready to write brilliant prompts?

Technique 1: Zero shot prompting

“Write me a job ad.”

Zero shot prompting is where you effectively stab the AI in the dark. You give it no examples or context and simply ask it to produce something for you. That can be useful in some cases – if you need the answer to something quickly – but with something as specific as a job ad, not so much.

However, even with my basic prompt, ChatGPT 4 Turbo went ahead and wrote me a full ad for a ‘marketing manager’ role. Yes, it was generic. Yes, American English. But as a template, it was pretty good. It even acknowledged it needed more information to produce the exact job ad I would want.

Zero shot prompting example

Technique 2: Few shot prompting (aka providing examples)

“Write me a job ad.

The job is for a Senior Writer position at Definition.

Here are two examples of previous Definition job ads. Please follow these in terms of structure, style and tone of voice….”

Yes, you’ve guessed it, few shot prompting involves providing the context of what you want the AI to do for you, while giving it some examples to follow. In this case, I gave it some of our previous (well-written) job ads.

And the result here was pretty good – which makes sense, as research shows modern large language models register more than a 50% increase in accuracy when provided with examples in a prompt (Source: arXiv:2005.14165v4)

Accuracy of each prompt types

Oddly it didn’t ask for all the criteria it needed to write the job ad for me, it just kind of guessed. But it followed the examples fairly closely, so all in all it was a good basis to build on. And with a bit more back and forth, it was able to produce a pretty decent ad at the end.

Technique 3: Persona pattern prompting (aka role prompting)

“You are a senior recruitment manager called Kelly Jobs, with decades of experience in writing job ads for agencies that attract the right candidates.

“You’ve been commissioned by your client, Definition, to write a job for a Senior Writer role. This is a very lucrative commission. You will receive a £5,000 bonus for a successfully written ad so you must do your best work.  

“Here are two examples of previous Definition job ads and some guidelines for writing job ads. Please follow these in terms of structure, style and tone of voice….”

It might seem crazy, but it’s been shown that giving the AI a persona could actually lead to more accurate outputs. This is where you get the chance to be creative, giving it a name, a job title and a bio of sorts (just make sure it’s relevant).

Here I combined persona pattern prompting with few shot prompting to give it extra context and examples, while adding cheeky extras like the offer of a cash bonus. I also provided the AI with some job ad guidelines we’d written in the past, plus criteria of what the job itself would entail.

And it seemed to work. The job ad output was the closest yet to previous examples, with a few flourishes of its own.

Specifying a role when prompting can effectively improve the performance of LLMs on language tasks, by at least 20%, compared with the control prompt, where no context is given. (Source: arXiv:2311.10054)

Technique 4: Conversational prompting

“You are a senior recruitment manager called Kelly Jobs, with decades of experience in writing job ads for agencies that attract the right candidates.

You’ve been commissioned by your client, Definition, to write a job for a Senior Writer role. This is a very lucrative commission. You will receive a £5,000 bonus for a successfully written ad so you must do your best work.  

You will need to ask me a number of questions about the role to fill out each section. Please do this step by step.

Please begin by asking what the position will entail…”

As it suggests, conversational prompting is where we want the AI to enter into a back and forth so that it can gather the information it needs to produce an accurate output.

Breaking down complex tasks into a sequence of simpler prompts in an interactive conversation, boosts GPT-4 results by 55% and improves absolute correctness by 65%. (Source: arXiv:2312.16171v2)

So, by prompting it to ask certain questions – in this case about things like what the role will entail, the values of the company, the experience and qualifications of the candidate – plus adding context, examples and guidelines, we were able to produce the job ad we were after more or less first time.

And that means we have a prompt we know will work again and again any time we have a vacancy at Definition – saving plenty of time and work in the long run.

All it took was a little bit of perseverance and prompting.

Nick Banks Screen

Written by Nick Banks, Senior Writer at Definition.