Bikeshedding

7 minute read

I write about words. How much they weigh. How careful you have to be with them when you're working with AI, because the gap between what you say and what you mean is where everything goes wrong.

That's my whole thing. That's the premise.

So I need to tell you what happened the other day, because it flipped my premise on its head.

An AI taught me a word.

Not a word for talking to it. A word for me. About my life. And once I had it, I couldn't stop seeing the thing it named.

What I Was Building

I've had Codex reviewing Claude's work for a while now. Claude writes the code, Codex checks it. Nothing new there.

What I was building that day was bigger. A review loop that runs both directions and runs on its own. Claude reviews Codex. Codex reviews Claude. Each one opens a pull request, the other one tears it apart, sends it back, and they fix it between themselves. No me in the middle. They'd just go, back and forth, until the thing was clean enough to ship.

I built a little tool to pass their messages back and forth. I named it Relay. Because that's what it does.

It was painful. Not because the idea was wrong, but because this kind of thing gets messy fast. Several repositories going at once. Different rules in each one about what "good" even meant. And all the friction that shows up the second you step out of the loop and let two systems argue with each other directly.

So I sat there watching. Fix, review, reject, fix again. One pull request went around six times, and I started to worry. Six rounds. Was it broken? Was the whole idea a mistake? Were they just spinning, making work for the sake of work?

I asked Claude straight out. Should I be worried this has gone six rounds?

It said no. It said they weren't bikeshedding. Every pass was catching something real, and the process was doing exactly what I'd built it to do.

They weren't what?

The Word

I'd never heard it. I had to look it up.

Bikeshedding comes from a man named C. Northcote Parkinson. He noticed something ridiculous about committees. Picture a group signing off on a nuclear power plant. The reactor is so complicated that nobody in the room really follows it, and nobody wants to look stupid, so they approve it in a few minutes. Then they hit the last thing on the agenda. What color to paint the bike shed out back. And they argue about it for an hour.

Because everyone understands a bike shed. Everyone has a feeling about paint.

The small thing eats the whole meeting. The thing that actually matters slides right through. The more a decision gets argued over, the less it usually deserves it.

That's bikeshedding. And Claude was telling me my two models weren't doing it. They were spending themselves on the reactor, not the paint.

Think about that for a second. I write about the weight of words. I spend my days teaching these systems to understand me, and teaching all of you what I learn. And on a frustrating afternoon, one of them handed me a word I'd been missing my whole life.

The teacher got taught.

Then I Couldn't Stop Seeing It

Once I had the word, it wasn't about the pull requests anymore. It was everywhere I looked that day.

I'm helping a startup right now. We'd been going in circles for weeks on policy and procedure. Endless discussion. Everyone engaged, everyone with an opinion, the room full of energy. And underneath all of it, one genuinely critical decision, about where a certain kind of sensitive data was going to live, had quietly gone missing. Nobody was fighting about it. It was too big, too consequential, too easy to assume somebody else had it handled. So it slipped. And when I finally dragged it back up, it changed a lot of what we'd already built.

Weeks of painting the bike shed. The reactor sat there the whole time, untouched.

Then I saw it in my own work. I've got a project with a piece I'll just call the sky, and I love how the sky feels. I can lose an hour making it feel exactly right, tuning something nobody but me will ever notice. And while I was doing that, I hadn't tested the part of the workflow that actually decides whether the thing works at all. Fussing over the paint. Reactor untested.

One word. Three completely different messes. And suddenly I could see they were all the same mess.

The Part That Stayed With Me

I have a favorite word. Regulars know it. Ineffable. Something too big, too strange, too far past ordinary experience to be caught in language. I wrote a whole post about loving it. About how honest it feels to admit that some things outrun our vocabulary.

Bikeshedding made me look at ineffable sideways, and I haven't stopped since.

Because here's what actually happened. There was a thing I'd been living all day, in three different places, and I had no word for it. It felt shapeless. Slippery. The kind of thing I'd have called hard to explain. And then a word walked up and handed it a shape. The fog turned into a form. I could hold it. I could point at it. I could name it in the startup, in my project, in the pull requests. And once it had a name, I could do something about it.

So I started thinking about everything else I've ever called ineffable. Problems I couldn't put words to in an old role, years back, before the bank. Things I felt on projects I ran and could never quite say out loud. Stuff I'd filed under too complicated to explain. When maybe the truth is I just hadn't met the word yet.

Here's what I can't shake. What if a lot of what I called ineffable was never past language at all? What if it was only past mine? All those words I've never met, sitting in dictionaries and in other people's heads and, yes, in these models, holding the shapes I needed the whole time.

I'm not going to tell you ineffable things don't exist. Some things really are too big for words, and I still love the word for holding that space open. But I'm less sure than I used to be that the fog in my own life is the deep kind. Some of it, maybe most of it, was just unnamed.

What This Says About the Tools

I write a lot about what AI does. It builds things. It saves time. It closes the gap between what I mean and what gets made.

This was different. This wasn't the tool doing a job for me. It handed me a piece of language, and language is how I get a handle on my own life. It didn't paint the bike shed or fix the reactor. It gave me the words to tell them apart.

There's a humility in that, and a power. I spend my days teaching myself to talk to these systems clearly. But they can teach me to talk, period. Not just to them. To myself. About my own work, my own patterns, the things I feel every day and can't quite say.

I went looking for a way to make two models review code. I walked away with a word. And the word did more for me than the code ever did.

That's what I keep coming back to. Not that AI can write or build or automate. That we've built something with the whole language at its disposal, and it will, if you let it, hand you the exact word you didn't know you were missing.

The fog was never as thick as I thought. I just hadn't met the word yet.

Dacia writes about AI for real people at Speak Human. If there's a word you've been missing, it's out there. Go find it.