Why your AI needs an archetype before it needs a prompt
“A Jedi uses the Force for knowledge and defence, never for attack.”
— Yoda, The Empire Strikes Back
Yoda has always fascinated me. Wise, curious and quietly confident. The sort of character who knows considerably more than he lets on. He listens, observes and shares his knowledge when needed. Admittedly, his sentence structure could do with a content designer, but we'll let that slide.
For me, his quote about the Force is less about power and more about responsibility. Knowing what you can do is one thing. Understanding what you should do is something else entirely.
That distinction has become increasingly important since I started working in conversational AI at the beginning of 2026.
AI can explain, reassure, guide and help someone move forward. But it can also overreach. It can sound confidently wrong, push when it should pause, or become the mate you didn't ask for, the adviser it isn't allowed to be, or the salesperson wearing a very convincing helpful hat.
In financial services, those aren't just annoying personality traits. They're potential risks. And one of my biggest lessons has been that designing an AI experience shouldn't start with the prompt.
It needs to start much earlier.
I started by asking the wrong questions
When I first started working on Aurora, Moneybox's AI assistant, I did what probably comes naturally to most content designers. I started thinking about the words.
What should the opening message say? How should Aurora respond? What happens when it doesn't understand a question? How do we stop it sounding like it's read one fintech blog and now thinks it's Terry Smith?
These are perfectly reasonable questions. They were questions our product designers, product managers and I were all exploring. But looking back, we'd jumped ahead.
Aurora already existed as a concept, and Moneybox had an established brand voice. I wasn't there to reinvent either. My challenge was to understand how those foundations could work in an experience that was part static journey, part conversation and part system behaviour.
The customer could ask questions, explore their financial goals and receive responses shaped by their circumstances. And unlike a traditional journey, I couldn't simply write and review every possible response before it reached someone.
That changed my thinking.
I wasn't just designing what the AI would say. I was helping define how it would behave.
Which meant we needed to establish its identity first.
Finding The Guide
To explore this, I designed a series of workshops with colleagues across product, design, brand and marketing. I'll admit, it was slightly terrifying. As I was still finding my feet in conversational UX, and there I was, helping define the personality and behaviour of an AI experience —no pressure, then.
But something I've learnt throughout my career is that you don't always need to walk into a room with the answers. Sometimes your value as a designer comes from asking the questions that help everyone find them.
We explored Aurora almost as if it were a person. How would it communicate? What qualities would it have? How should customers feel after interacting with it? From there, we explored different archetypes.
A friend could feel warm, but become overfamiliar. A coach could be encouraging, but drift towards telling customers what to do. A mentor could offer wisdom, but risk feeling distant. The Guide gave us something different.
Our version of The Guide helps customers understand the financial landscape, explore their options and move forward with greater confidence. It explains without lecturing. Supports without pushing. Gives direction without making the decision. And that last point matters enormously.
In financial guidance, the customer must remain in control. Aurora can explain options and trade-offs, but it must not present personalised financial advice as guidance or imply that a particular product is the right choice.
That's not simply a tone of voice decision. It's a safety boundary.
Defining what Aurora isn't
One of the more interesting things I learnt was that defining what Aurora isn't became just as valuable as defining what it is.
We established a set of personality boundaries:
Helpful, but never vague.
Clever, but never smug.
Inquisitive, but never invasive.
Present, but never overbearing.
Human-like, but never fake.
At first glance, these could easily be mistaken for another set of brand guidelines. But they're more useful than that. They're design constraints.
Take "inquisitive, but never invasive". Aurora needs information to help customers explore their financial goals. But asking questions isn't enough. We need to explain why we're asking, what that information helps us understand and when we've gathered enough.
Or "human-like, but never fake". We want conversations to feel natural. But there's a difference between communicating with warmth and pretending the AI has human feelings or experiences. That distinction becomes particularly important when someone is worried about money or experiencing financial difficulties.
The system needs to respond appropriately, acknowledge its limitations and provide a route to further support when needed. It shouldn't manufacture reassurance it cannot justify.
Trust isn't created by making AI sound more human. It's created by making its behaviour more dependable.
The difficult bit came afterwards
With the archetype defined, I thought translating everything into a system prompt would be relatively straightforward.
A little lift and shift. I was wrong.
Telling a room full of designers that Aurora is "The Guide" works because we've collectively explored what that means. An LLM doesn't share those conversations or assumptions. It needs explicit instructions.
"Be calm" becomes "Use steady, direct language. Avoid exaggerated praise and explain what happened when something goes wrong."
"Keep the customer in control" becomes "Explain relevant options and their risks, make the limits of guidance clear and leave the decision with the customer."
"Be inquisitive" becomes "Ask only for information needed to answer the question, explain why it's needed and avoid unnecessary follow-up questions."
And even then, the work isn't finished.
A well-written system prompt doesn't guarantee safe, consistent behaviour. You still need technical safeguards, clear boundaries, testing and ongoing evaluation.
The archetype gives everyone a shared understanding of what we're trying to build. The system prompt begins to translate that understanding into behaviour.
That was a significant shift in how I understood my role.
We're designing more than words
Looking back, I didn't appreciate just how much of this work would involve designing the system rather than the content appearing on screen.
I'm still applying the principles I've used throughout my career. Clarity, accessibility, user needs, structure and trust. But I'm applying them differently.
I'm helping define relationships, behaviours, constraints and the decisions an AI should never make on someone's behalf. And that raises an important question about where content design is heading.
If we're not involved in defining how AI behaves, those decisions will still happen. They'll simply be shaped by technical defaults, commercial priorities or vague instructions to "make it sound friendly". That's not a future I'm particularly keen on.
The role of content design isn't getting smaller. The room is getting bigger. And yes, we'll probably still spend far too long debating button labels. But we're also helping shape systems that can influence how people understand information, explore their options and make decisions.
Before your AI needs a prompt, it needs a role. Before it speaks, it needs boundaries. And before it guides someone, we need to understand what kind of guide we're building. Because a good guide doesn't need to have every answer.
It needs to know when to help, when to question and when to step aside.
