What Makes a Good AI Prompt?

Most verdicts on AI are reached in under a minute. Someone types a sentence into a chat box, reads a mediocre answer, and concludes the technology isn’t ready, summarily dismissing the potential value of AI tools. I’ve watched it happen in real time and I’ve had direct conversations where that story has played out. 

The problem is that bad prompting rarely registers as a problem at an organisational level. There is no metric tracking effective prompts. But people feel it, and that ultimately affects how AI is used across the organisation. A poor first experience like that chips away at their trust in the whole idea, which can explain a lot of the resistance and scepticism companies see and why so many AI tools sit at a fraction of their potential.

What a good prompt contains

Our knowledge of prompting has evolved and there are many frameworks out there, but strip them back and they’re all encouraging the same information to be provided: the role you want it to play, who it’s for, what you’re trying to achieve, the constraints, examples of good and bad, and what ‘done’ looks like.

The same information you would provide in a brief to a human being to get the job done right. Why would it be any different for an AI?

AI isn’t magic, and it’s not a mind reader. So for my mental model, the most useful prompting habit isn’t a secret phrase or a framework. 

It’s empathy. 

Put yourself in the model’s position: what does it need to know in order to give me what I want? What would a new hire need to know to do their job well? And then act accordingly. 

With that thinking, it quickly becomes obvious that a one-line prompt is almost never enough and pages are too much. And if you find yourself pasting pages of background into every prompt, that isn’t a prompting problem at all. That’s a context problem, and we’ll come to it next.

Let the model help you ask

One of my favourite techniques is also the laziest one: ask the AI to help you write the prompt. Describe the task roughly, tell it to ask you whatever it needs to know before starting, and answer its questions. Models are surprisingly good at spotting the ambiguity you didn’t notice in your own request. If there’s one thing you take away from this article, make it this one. Try it out and see if you get better results.

Write the brief once

Inside Vaul, anything that should be done the same way, re-prompted in the same way, regardless of who’s doing it, gets written down once as a reusable brief, and improved every time a piece of work teaches us something new. The tooling calls these ‘skills’. The clearest example is how we run our engagements. Regardless of which consultant or team works on a client, the output should be the same. The way we write up a call, structure a report or prepare a workshop stays the same, because that consistency is part of our brand. A skill is that expectation written down. Nobody has to hope that everyone prompts well on the day, and when someone finds a better way of doing something, it goes back into the brief for everyone rather than staying in one person’s head.

Which is why I keep coming back to training and enablement. People start trusting AI the first time they watch it genuinely improve their own day, whether it’s shown in quality or productivity or innovation, but they can’t get to that moment while the basics of prompting are still tripping them up. Helping teams over that hump is a big part of what we do at Vaul, and personally, it’s some of the most satisfying work: watching someone go from sceptic to enthusiast. They’ve seen the value for themselves and they want more of it.

There’s one thing a better prompt won’t fix, though. We’ve watched the same, perfectly reasonable prompt produce generic filler in one system and work you could put in front of customers in another, without a word of it changing. The difference was never the prompt. It was what the system knew before it started, and that’s where we’re going next: context.

Part 3: Why Context Matters in AI