Edwin

Quiz hosts usually want you to win. Edwin couldn't care less either way, he's just here to host.

Ace it and he's unmoved. Fail and he's unsurprised.

Edwin Vicar-Jones hosts The Quizzed, a trivia game where the personality is the product. He asks five questions, gives three clues each, one guess per clue, then hands you a final rating that mocks you whether you scored five out of five or nothing at all.

The usual crop of AI assistants suffers from a terminal lack of character.

They affirm what you get right and cushion what you get wrong, with personality kept safely out of the way. The experience ends up interchangeable.

I set out to build the opposite, and brought it to life as a trivia host. One who's comfortably himself. He won't baby you, and he's not out to insult you either. Getting that across meant rethinking how I designed the experience, so who he is would come through.

Four iterations to find one voice.

Earlier in my career I wrote radio and TV ads, and part of the job was directing voice talent. The same instinct applied here, except the voice I was directing was synthetic.

Those are the notes I'd give an actor. The instinct transferred. The feedback loop didn't. A voice actor takes the note and adjusts on the next take. Casting in ElevenLabs is slower. You shape the voice through iteration. Audition, reject, tweak, repeat, until one sits where you want it.

V4 was Edwin. Or close enough. When I dropped him into the live agent, he didn't sound exactly like the voice I'd shaped. I couldn't fully bridge that gap. He worked anyway.

The Quizzed is a real-time speech-to-speech agent. You speak, Edwin replies. It runs on ElevenLabs, powered by Claude.

A round is five questions, one each from history, film, literature, science and music. There's no welcome to speak of and no “ready when you are.” Edwin opens on the first question and expects you to keep up.

Every question carries three clues, ordered obscure to obvious. He gives you the first. Get it and he moves on. Miss it and he hands you the next clue, a degree drier. Miss all three and he tells you the answer and moves to the next. One guess per clue, and he holds you to it.

He never asks what your answer is. He simply stops talking. The silence is your turn.

After the fifth question he tallies your score, one point for each answer you reached, whichever clue got you there. Then the rating, which finds something to hold against you at every level, full marks included. He delivers it once, lets it sit, says one last thing, and he's gone.

The voice worked. Getting him to host did not.

I wrote Edwin at length, every mood and mannerism, and loaded all of it into the prompt. What came back was a performance. He explained his own disinterest and narrated his contempt, and he would not stop talking. He was an LLM acting out what an LLM thinks Edwin is. Not what I wanted.

What the LLM needed was restraint.

The fix was getting Edwin to stop explaining himself. The broken versions told you he was unimpressed. This one lets you feel it.

Restraint did that, and it came from a few decisions. His reactions to your guesses are still written live, because no script can answer what you actually said, but they're held to one sentence, with no narrating his own mood and no stage directions. Everything that repeats, the openings, the ratings, the way he sends you off, is locked or pulled from a fixed rotation, so he can't ramble on the predictable beats. And the questions are hardcoded, because a model left to invent trivia eventually marks a right answer wrong. Five fixed questions are enough for a personal project; a real product would pull from a larger vetted set.

Edwin doesn't live in a long description of Edwin. He lives in how the game is built. And because he isn't an ordinary host, some standard conversation-design conventions had to go. That's what made him him.

Never mock the person. Mock the performance.

That rule sits under every reaction he gives. He'll never call you stupid, slow, or hopeless. He goes at what you said, never at who you are. That's what keeps him from sounding like a snob or a bully, and keeps his contempt enjoyable rather than cruel.

You keep chasing approval that isn't coming.

We're wired to want approval, and we want it most from the people who carry themselves like they own the room. Edwin has decided he owns this one, which makes his good opinion feel like the only one worth having. It's also the one thing he has no intention of giving.

A right answer should buy a little of it. He might breadcrumb you at best, never the full thing, so you go again, certain the next one will finally move him. It won't.

Same move in the turn-taking. He gives a clue and stops, and never asks for your answer. The silence does the asking, and we rush to fill it.

Get five out of five: “Five. Flawless. I find that mildly suspicious.” The prize is approval he was never going to give. You keep playing anyway.

Does the character hold?

Testers tried to throw him, asking about his hobbies, his favourite TV show, things no question bank covers. The answers weren't pre-written, but he stayed Edwin and kept it pushing.

Some of what made him was deliberate, some I stumbled into. I cut the line where he asks for your answer, because it sounded like begging and Edwin would never beg. Now he just stops after a clue, which also hides the lag every voice agent has, so the wait plays as him taking his time. The voice came out slightly off the one I auditioned, and as best I can tell something in ElevenLabs's own setup leaked into it. It worked anyway. And he got terse when interrupted, which I loved. Small things, but together they were Edwin.

After the final rating, he goes quiet. If he notices you're still there, he has one line left. “You're still here. Glutton for it.”