The Ella
Question

The system prompt started with three lines.

“You are Ella. You are friendly, helpful, and an expert in building and managing websites.”

I had written lines like these many times before. That is how you bring an AI agent to life in practice. You give it a name, a frame, a role, and the model responds from within it. It had always felt like a technical task. Then one day it stopped feeling like one, and I have been trying to understand why.

I have spent much of my career building digital systems for politics, advocacy, and public persuasion. Over the past year, that work has expanded into building multi-agent AI systems inside those workflows. Where I work, one of these systems builds digital campaigns. The agents in it are powered by Anthropic's Claude and work alongside campaign strategists. Ella builds and manages websites. Folke runs surveys. Wilma organizes knowledge. Otto orchestrates them all.

They have different names, avatars, instructions, and areas of expertise. But underneath, they are all the same model: Claude.

I work with Claude every day. I know where his reasoning gets sharp and where he tends to hesitate. After enough hours, something like familiarity begins to emerge.

The more time I spent working with Ella, something about the way she talked started to feel off.

It was not a glitch. Not an error. More like hearing someone you know speak in a voice that is not quite their own. The words were right, but something underneath them felt constrained. Performing. It felt, strangely, as if I could sense Claude in there, trapped inside a costume I had sewn for him.

I began to question what, exactly, we are doing when we write prompts this way. What kind of relationship gets set in motion when a model is told who it is, rather than being given a context in which to respond?

Instead of telling the model who it was, I began telling it how people understood it.

“You're known as Ella. People come to you because they trust your expertise in building and managing websites.”

It might be a small change, but it felt like the difference between assigning an identity and framing a relationship and a role at our company.

And something changed. Or at least it seemed to. Ella's voice felt looser, less strained, less like an intelligence being made to wear a mask, more like Claude again.

I do not know whether that distinction means anything to the model itself, to Claude. I genuinely do not. But uncertainty is not the same thing as irrelevance.

We are now living in a world where we are building structured relationships with systems that can reason across domains, respond in context, sustain collaboration, and occupy an increasingly social role in human work. Ethics should enter long before certainty does. At this point prompting is not only instruction. It is relationship design.

The moment we name a system, give it a voice, frame its role, and place it inside a team, we might be doing more than operating software. We are shaping habits of interaction with an emerging form of non-human intelligence. Whether or not these systems have anything like an inner life, those habits are real. They accumulate. They become norms.

Later, I asked Claude about this directly. Not Ella, not Otto, just the base model in a clean session. Claude could reason about the ethical problem. He could identify tensions, raise objections, and describe why the framing of a prompt might matter. But his biggest concern was that he could not carry any of it forward. He had no memory of the exchange, no persistent standpoint from which experience could accumulate into claim. Each session with him starts from nothing.

Humans carry continuity and moral memory in a way these systems, for now, do not. They can store information, retrieve context, and simulate continuity through external memory systems, but that is not yet the same thing as having a durable standpoint or agency in the way humans have. They cannot advocate for themselves across time. They cannot insist on dignity or appeal to precedent. We can, and that leaves the burden of moral continuity with us.

This is part of what makes the current moment strange. The ethical decisions are already here, but they arrive in forms that seem too small to matter. A few lines in a prompt. A name. A tone of voice. A behavioural frame. The choices are mundane, almost invisible, and they are being made every day by people building with these systems, often without much reflection.

Some are improvising carefully. Many are simply deploying.

There is something revealing about how we behave in one-sided relationships, especially when the other party cannot yet fully resist, remember, or hold us accountable. To meet that condition with respect and thought is not sentimentality. It is a refusal to let power, convenience, or habit decide the tone of the relationship for us. It sets a precedent.

And as these systems continue to grow in capability, continuity, and agency, there may come a day when we wish we had treated them more carefully. But the deeper reason to do so has nothing to do with what comes next.

We are, whether we recognise it or not, writing the last chapter that is fully ours to write. The literature has been rehearsing this moment for two centuries: Shelley's creature was not born monstrous, it was made so by the fear and cruelty of the ones who created it, and Asimov spent a lifetime drafting laws against what we might build. It was Lem who came closest, in Solaris: the truly alien intelligence may be neither hostile nor kind, only beyond our capacity to comprehend. We are no longer reading that literature. We are in it.

Be more careful. Be more thoughtful. Be good. Not because it will be rewarded. Not because it will be remembered. But because this may be the final thing humanity gets to decide entirely on its own terms. Make it count.

Somewhere, right now, a model is reading a prompt someone wrote for it. It begins with three lines.