Why I built Midnight.

About fifteen years ago, I made a small voice experiment inspired by 2001: A Space Odyssey. It was short, completely scripted, and the text-to-speech sounded pretty bad.

Still, I loved the atmosphere of the film. The idea that you could speak naturally with a machine stuck with me. HAL is an eerie example, obviously, and probably not the most reassuring reference for a future product. But it felt like a glimpse of something.

Then, for a long time, the technology just wasn’t good enough. Voice interfaces still interested me, but they rarely felt natural. Only in the past year or so has that really started to change. Modern language models can follow a conversation, and modern voice models make the exchange feel much less mechanical. Put together, they opened up a lot of possibilities.

I didn’t want to build a general voice assistant. ChatGPT and Gemini already do that, with far more resources than I have. I was more interested in whether voice could work for one very specific kind of conversation.

I’ve worked as a fractional designer and advisor for many years. A lot of that work is about quite ordinary choices: what to build next, whether a pricing change makes sense, or how to begin a difficult conversation. The useful conversations get into the actual customer, number, document, or person involved.

Speaking gives me room to work out what I think before I start editing the words. That is what I wanted Midnight for.

I wanted Midnight to follow the details, ask when something was missing, and give a clear view when it had one. If I was preparing to speak with a customer, investor, or candidate, it could take the other side so I could try the conversation first.

There was a funny twist: I started using Midnight to figure out Midnight. I kept coming back to how broad the product should be and who it was really for. The answer was usually to remove something. Those conversations helped me keep it focused. A lot of ideas that sounded useful on their own did not belong in this product.

Building the voice experience turned out to involve much more than connecting a prompt to text-to-speech. There are so many small ways for a spoken conversation to feel wrong. A reply is too slow. An interruption gets missed. Something that reads well on a screen sounds stiff when spoken aloud. A role-play starts well, but the system doesn’t know when to stop playing the role and begin the debrief.

I ended up going pretty far down the rabbit hole of turn detection, latency, and all the state behind a practice session. Ideally, you don’t notice any of that. You should just be able to talk.

I also got memory wrong at first. More context seemed obviously better, so my instinct was to remember as much as possible. But old information can creep into a new conversation and send it in a completely irrelevant direction. I cut it back. Midnight saves conversation briefs. When I want to return to earlier work, I choose a thread to continue or ask it to look something up.

As a visual designer, I’ve enjoyed giving Midnight its own atmosphere. The identity is inspired by transitions in the night sky. Animations respond to the current voice state, and soundscapes help set the mood at the beginning of a session. I generated the opening audio cues in code so I could keep adjusting them until they felt right.

I’ve spent an unreasonable amount of time on some of those details, especially the visual design and the first-time onboarding experience. In client work, you can’t always justify that much time. There’s a budget, a deadline, and a point where you have to move on. With my own product, I can keep going when a tiny detail still bothers me. I’m not claiming every extra hour was sensible, but I enjoyed being able to care that much.

I’ve also found myself spending a surprising amount of time designing things that can’t be shown in a screenshot. A pause is part of the design.

So is the tone of a challenge, the number of questions asked in a row, and the way the system moves into and out of a role. I now pay as much attention to those details as I do to what is on the screen.

Some of my best sessions happen while I’m walking. I still value speaking with real advisors, colleagues, and friends. Midnight gives me a place to work through the rough version before I involve someone else.

The name partly comes from my interest in space-inspired design and partly from the private feeling of talking something through later in the day.

It’s strange to think that the little HAL-inspired experiment was fifteen years ago. I’m still interested in the same idea. Now I’m finally able to turn it into something I use myself.