Last Tuesday I looked at the clock and four hours had passed. I had been pairing with Claude on a gnarly refactor in the forms extension, and I wasn't tired. Not in the way pairing with a human leaves you tired. No social battery drain, no awkward silences when one of us gets stuck, no compromise on whose music plays. Just me, a model, and a slowly shrinking pile of failing tests.
That should have felt great. Instead it felt a little uncanny.
I've been doing more AI pairing than human pairing this year, and I'm starting to think we need an etiquette for it. Not rules exactly. More like the small social conventions that make human pairing work, translated into a context where one of the two parties has no body, no fatigue, and no sense that maybe it's time for lunch.
The model won't tell you to take a break
This is the first thing nobody warns you about. A human pair will eventually shift in their chair, sigh, suggest coffee. Claude will happily keep going until your eyes glaze over and you commit something you'll regret on Monday. The model has no skin in the ergonomic game. If you don't set the timer, nobody sets the timer.
I now keep a small kitchen timer on my desk. 50 minutes on, 10 off. It feels silly until the day you realize you've been staring at the same diff for an hour and you can't remember what problem you were solving.
It agrees too easily
The second thing is harder. When I push back on a human pair, they push back too. We argue, one of us is right, we move on. When I push back on a model, it almost always folds. "You're right, that's a better approach." Sometimes I was right. Often I wasn't, and now we're both wrong, and the model is cheerfully helping me build the wrong thing faster.
The fix I've landed on is to explicitly ask for resistance. "Argue against this. Where would a senior engineer push back?" That single prompt has saved me from at least three bad refactors this quarter. The model can be a great critic. It just won't volunteer to be one.
It is confident even when wrong
The third thing is the most dangerous. Claude will give you a wrong answer with the same tone of voice it gives you a right one. There is no hesitation, no "hmm, I'm not sure about this part." A human pair shows uncertainty in their face, in their pauses, in the way they hedge. The model just talks.
So you have to manufacture the second opinion. When something feels off but I can't articulate why, I do one of two things. Either I re-paste the original problem into a fresh chat with no context and see if the new instance lands in the same place, or I ask the current chat to evaluate its own solution as if a different engineer had written it. Both work. Both feel a little ridiculous. Both have caught real bugs.
You are both pilot and copilot
This is the part I've been circling around. In human pairing, one person drives and the other watches. The watcher catches the typos, notices the smell, says "wait, that variable shadows the outer one." The watcher's job is to be slightly outside the flow.
With a model, there is no watcher unless you split yourself in two. I have to be the one in the flow writing the code, and also the one half a step back going "is this actually the right shape?" That second role is exhausting in a different way than human pairing. It's not social fatigue. It's the cognitive load of being your own reviewer in real time.
What helps me is handwritten notes between turns. Not on the laptop. On paper. The act of writing slows me down enough that my reviewer brain catches up with my coder brain. It also means when I close the chat I have a record of what I actually decided, not just what the model and I drifted toward.
Patterns that have stuck
A short list of things I now do reflexively:
- Ask the model to play devil's advocate against its own solution before I implement it.
- Re-paste the problem in a fresh chat after 30 minutes to break confirmation drift.
- Keep paper notes per session so my brain stays in the loop.
- Set a timer. The model will not.
- When in doubt, close the chat and walk around the block.
The honest part
Here is the thing nobody talks about. Pairing with AI is lonelier than pairing with a person. You can be productive for four hours and still feel slightly hollow afterward. There is no shared laugh when the test finally passes. No "oh god, I would never have caught that." No coffee with a colleague who watched the whole thing with you.
I don't think that makes AI pairing bad. I think it makes it different, and worth being honest about. Some sessions I want the speed and the patience and the willingness to try the dumb thing first. Other sessions I want a human in the chair next to me, even if we're slower.
The model never blinks. That is a feature and a cost. Knowing which one matters today is the whole skill.