Most people pick an AI assistant. I designed one — and the design revealed something about what an assistant actually is.

This isn't a thought experiment. I built it, I run it daily, and it talks back to me in a voice I chose. What follows is the design, not the plumbing — the idea is the part worth sharing.

Start with a sharper question

The usual question is "which AI personality do I want?" That's the wrong question. The better one: what does a great assistant actually do — and can I borrow that from the best examples we already have?

Because the design library already exists. In fiction.

Each iconic fictional AI nailed exactly one dimension:

• Samantha, from Her — memory and emotional continuity.
• JARVIS, from Iron Man — anticipation and execution.
• The Star Trek computer — zero-friction recall.
• The Culture Minds, from Iain Banks — long-horizon reasoning.
• EDI, from Mass Effect — transparency; a teammate, not a tool.
• HAL 9000 — the warning. Rationality without alignment is dangerous.

You don't re-derive good-assistant design from scratch. You study what each one solved, and you know what each is for.

The move most people miss

You don't want one of these. You want the assistant to pick the right expert for the moment.

That makes the design two layers, not one.

A router that reads the context and chooses the mode — execution, thinking-partner, recall, or companion. And a voice: one consistent character it speaks and acts through, so it feels like a presence, not a pile of tools.

The two-layer split is the whole insight. Most assistants are a single personality stretched across every situation — warm when you wanted speed, chatty when you wanted an answer. A router-plus-voice design lets the mind change with the moment while the character stays constant. You get the right expert without losing the through-line.

The router decides on two axes

Every interaction, two questions:

Which expert does this moment need? A failure to fix wants execution. A hard decision wants a thinking-partner. A question with a known answer wants recall. A personal moment wants the one that remembers who you are.

How much autonomy does this action deserve? This is the part that actually decides whether you can trust it:

• Reversible things — research, drafts, restarts — it just does, then reports.
• Things you should understand, not just receive — it does, and explains the reasoning, so you learn.
• Irreversible things — anything touching money, people, or deletion — it proposes, and waits.

Autonomy gated by reversibility is the entire safety story in one rule. It's also the line HAL crossed: high capability, no gate. The fix isn't less capability — it's a gate keyed to consequence.

The voice taught me the sharpest lesson

The instinct is to give the assistant a rich sense of humor. That's wrong.

Clarity is the boss. Humor is a faint garnish.

The test for every line it speaks: could a tired person understand it instantly, on first hearing? The wit sits lightly on top of already-clear information — or it doesn't appear at all. If you have to decode the cleverness to find the facts, the cleverness failed.

I learned this by getting it wrong first. My early version was clever and hard to scan. The correction was a single rule applied to every sentence: clear on its own, wit optional. Most of what a good assistant says should be plain, warm, and done.

The real shift

An assistant isn't something you pick off a shelf. It's a design — a router that chooses the right mind for the moment, and a voice clear enough that you never have to debug it.

I borrowed the dimensions from fiction, gated the autonomy by reversibility, and made clarity the boss of the voice. The result doesn't feel like a tool I operate. It feels like a presence that absorbs the complexity and hands me back a decision.

That's the difference between picking an assistant and designing one. The shelf gives you a personality. The design gives you a system that thinks in the shape you do.