Prompt like nobody's watching

My favourite thing about using AI is that it’s endlessly patient with my nonsense and my persistent skipping of pleasantries. Often it’s just that I’m rushing and I know I can’t hurt its feelings. Sometimes, I’m impatient because the first response I get isn’t quite what I wanted. For whatever reason, my most-used prompt when creating content with a GPT is, “Not that - try again.”

Other times, I’m more affable but shockingly unstructured. Scrolling back through past threads, it’s as if I think full stops are for wimps. For example: “Hey, I need to work out a plan for an activity day with some people who do [some loosely-described but very important jobs] and they need to learn about [some poorly explained but still real things] so they can collaborate and be better at [doing some vital stuff] and we’ve got maybe 6 hours with some decent breaks and they’re at cabaret tables I think.”

It's like a toddler's stream of consciousness. But a toddler with a real, professional job.

Incredibly, the AI comes back with a nicely drafted, well-structured piece of work, my AI and I remain on excellent terms, and we go on to have a pleasing back-and-forth to refine The Thing until I’m happy with the tone and content, and the AI is happy with the spelling and grammar.

It feels miraculous every time.

The point is, there are two things your AI isn't:

1. It's not human. I almost never write a prompt that looks the same as a brief that I would give to a person. The AI doesn’t need a meeting to warm up to the idea of helping out. I’m not worried about what it thinks of me or whether it’s totally understood the point of it all; I’m not worried it will get fed up or de-motivated. It's the most transactional relationship in my life - not counting my relationship with my cat. And that saves me time and emotional labour I can devote to people in my life who do need that. It is, though, a shockingly amenable and personable assistant.

2. It's not Google. The other trick we’re all learning when we prompt natural language AI tools is to de-programme ourselves from everything we’ve learned about getting the best from a search engine. Specifically, we have to forget our well-honed Google search skills. If you type in 'HR policy' it probably wants to say, 'Yeah - what about it, Boomer?'* Fortunately these things are pretty well trained so it'll try and be more helpful and work out what you want. So help it out.

At this point, form dictates that I offer some tips. So here’s mine:

Play with it. Everything I know about prompting is from trial and error, advice from colleagues, or things I read online. Try stuff out. You didn't go to Googling classes but you nailed that skill. You got this.

Be clear and concise. Or not. The AI is best with clear instructions, but ‘clear’ means something slightly different when you don’t need to worry too much about having the perfect word or perfect punctuation. Tell it what you want in the same way you’d tell your colleague if you both had no imposter syndrome, self-consciousness or need for approval. It doesn't care if you repeat yourself.

Give context. I know I said you don’t have to talk to it like a person, but maybe think of it as the new hire who hasn’t learned your very niche jargon and missed the first 20 meetings. Give it stuff to go on so it understands your intent.

Iterate. Fine-tune. “Make it warmer.” “Make it shorter.” “Make it punchier.” “Talk about the cat more...” You get the idea.

Use examples. Or even metaphors. If generative AI is anything, it's a bad poet. Try some ‘like’ or ‘as’ exposition. Try asking it for exposition. If you don’t have the words, think how else you can convey the scenario.

Stay curious. Ask it questions, try new releases to see if you get different responses. Our founder will tell you that if he takes two weeks off the AI news feeds, he feels like a historian. He’s not wrong.

Be yourself. If you're more comfortable writing in another language, do that. If you want it to always write in the manner of Taylor Swift lyrics, tell it.

‍Know that it’s a lazy bu***r. AI is good, but it’s like the smart kid at the back of the class who already worked out the teacher’s answers were on the last page of the book. Ask a GPT to show you examples of something, and it gets bored after reading about 1,000 words and giving you 20 examples. Ask for more. Ask again. For the love of Brian, check the citations are real. Which brings me on to...

Don’t trust it. A lot of the free online tools you’re using probably already own your data. Don’t get caught out entering information that becomes part of someone else’s training model. Forever. And I’m not just saying that because we make secure, private versions, although it helps.

Seriously. Don’t trust it. AI will lie to your face and be every bit as compelling as when it finds you true information and makes nice things. It won’t be sorry. Check your sources before you rely on them for something that really matters.

*It really doesn't think that, so don't worry. In fact it doesn't judge you at all. That's why you don't have to worry about asking silly questions.