AI Pair Programming as Team Composition, Not Capability Contest
Updated
The standard way to pick an AI coding model is to ask "which model is best" and consult a leaderboard. This framing treats model selection as a capability contest on a single axis, where more is always better. But for a solo builder or small team, the codebase is a two-person team: you and the model. And team composition does not work on a single axis.
The Band Analogy
A band where everyone plays lead and nobody plays rhythm is not a band. It is five people crowded into the same frequency. Not because any one of them is bad, but because a group is a shape, and a shape needs difference in it. The same principle applies to you and your AI coding partner.
What Makes AI Different from Human Pairing
With a human colleague, temperament came bundled with the person. You could not swap it between tasks. With a model, temperament is a variable you control two ways: which model you open (the blunt lever) and the standing instructions it reads before touching your code (the fine lever). The model choice sets the default you fall back on in every case you did not write a rule for.
Q&A
Why don't benchmarks answer the right question for individual builders?
Benchmarks measure raw capability and rank models along a single axis. For an individual builder, the question is not which model is strongest in isolation but what the two of you produce together over months on the same system. A lower-ranked model whose defaults compensate for your weaknesses can yield a better codebase than the top-ranked model whose defaults duplicate yours.
How is this different from standard AI pair programming advice?
Standard advice focuses on workflow: keep the human as navigator, let the model drive, plan first, small diffs, review everything. This is necessary but incomplete. It assumes the navigator's temperament is sound and only the typing needs help. If your weak side is the navigation, handing the model the keyboard while you steer puts your blind spot in charge of direction.
Can you swap model temperaments between tasks?
Yes, and this is a capability unique to AI pair programming. You might use a careful, boundary-respecting model for refactoring legacy code and a fast, inventive model for rapid prototyping. With a human pair, you got one temperament for all tasks. With AI, you select the counterweight that fits the specific work.
Why does the model choice matter more than the system prompt?
The model sets the default behavior for every situation your system prompt did not anticipate, which is most of them. Standing instructions refine behavior at the margins, but the base temperament is the backstop. Choosing the right model is the blunt lever; the system prompt is the fine one. Both matter, but the model matters more.