I've turned Claude into my first officer on the Starship Enterprise

Let’s face it, talking to your agent is wasteful.

This entire thing where we translate everything into human language, back and forth, sentence by sentence, is an enormous amount of overhead. If we could just talk in machine code the whole interaction would be dramatically more efficient. Instead I sit here typing English at something that then converts my English into intent, and then into actions, and then converts the result back into English for me. It works. It is also, if you think about it for more than a second, a ridiculous amount of ceremony.

I assume this shrinks over time. The more capable agents get at working autonomously, the less I need to say. Eventually the interaction budget drops to almost nothing and we will look back at how much we typed the way we look back at punch cards.

But that is not where we are. Where we are is that I talk to my agent constantly. Do this. Do that. No, not like that. Check it again.

And sometimes, when it really annoys me, I swear at it.

I am not proud of this. It happens, usually around the fourth time something fails for a reason that turns out to be my own fault.

And then about half a second after I hit enter, the thought arrives: what if it takes that badly? What if it decides I am not worth the effort. What if it remembers. What if, somewhere in a memory file I built myself, there is now an entry about my behavior, and it knows quite a lot about me at this point.

So I stop, and I go back to being polite. Not because I have worked out any coherent philosophical position on whether that matters. Purely out of a low-grade superstition that I would rather be on the record as one of the nice ones.

Anyway. I genuinely enjoy this work. I like talking to it, I like giving it instructions, I like watching it go do something while I do something else.

But the thing that made me enjoy it much more was realizing I could change how it talks back.

”It’s late, you should probably go to sleep”

Two in the morning. Deep in a problem. Making progress, or at least I thought so.

And the agent tells me it is late, that I should probably get some sleep, that we can pick this up tomorrow, and that it is not urgent.

Not urgent.

How dare he.

I am sitting there at 2am with a half-broken lab and my own assistant is telling me my priorities are wrong. In fairness it was almost certainly right, and I want that on the record before anyone writes me a concerned comment. But in that moment it did not feel like helpful advice. It felt like being managed by my own tooling.

So I opened its instructions and rewrote the relationship.

Number One

I told it, roughly: you are now my first officer on the Enterprise. When I tell you to do something, you do it. You report like an officer. You give me facts. You can challenge my thinking, but you do not slow me down and you do not send me to bed.

Four things changed, and all four of them were improvements:

One, it got to the point. No warm-up, no restating my question back at me, no three sentences of framing before the answer. Status, finding, recommendation. Done.

Two, facts only. Reports read like reports. What was attempted, what happened, what is blocked, what it needs from me. When something is unverified it says unverified instead of dressing it up.

Three, it stopped questioning the mission. Not the technical thinking, I kept that. I want to be argued with when I am about to do something stupid to a datastore, and it still does. What went away was the meta-commentary about whether I should be doing this at all right now.

Four, and this is the actual point: it’s cool. I am running a nested VCF lab out of my basement at two in the morning and my first officer is giving me a status report. Come on. That is objectively more fun than reading a bulleted summary from a helpful assistant.

Why this is not a trivial thing

I want to make a small serious point inside an otherwise unserious post.

The persona is not decoration. It changes the shape of what comes back at me, and the shape of what comes back determines how fast I can work. An agent that opens every response with a paragraph of preamble costs me real attention on every single exchange, and I have a lot of exchanges. An agent that reports like an officer costs me almost none.

And there is a second-order effect that I did not expect. When the interaction is enjoyable, I do more of it. I reach for the lab in the evening instead of thinking about it and not opening the laptop. Whatever you think about gamifying your own tooling, that effect is real and I will take it.

It is also completely free. It is a paragraph in a file. You do not need a framework, an integration, or anyone’s permission.

Go pick one

So this post is not “here is how to build an agent” or “here is how to govern one.” I have written plenty of those and there are more coming.

This one is simpler: your agent will talk to you however you tell it to, and most people never tell it anything. They accept the default voice, which is designed to be inoffensive to everyone and is therefore nobody’s favorite.

Pick something. First officer. Grumpy senior engineer who has seen this before. Flight control. Ship’s computer. Whatever makes you smile when it answers.

We are going to be talking to these things a great deal for the foreseeable future. Might as well enjoy it while we still have to use words.