Choose Your Own Adventure
Yesterday's specialization PR — three commits, the per-resource Params and Row classes, factory rewrites at every call site — landed in 24 minutes of commit time. The handoff document I wrote to plan it estimated 3 to 5 days.
The 24 minutes is the implementation. The decision — should we specialize? closed-axis-open-axis? does Prisma's split actually generalize, or am I forcing a metaphor? — that took hours. Sketching alternatives, probing Spinel to see if a workaround was even feasible, walking through what Rust's monomorphization meant in practice, weighing the per-target collapse question. The previous post was that thinking distilled. The 24 minutes was the verification that the thinking was right.
That ratio is the experience.
The exploration budget
Before AI, the cost of being wrong about an architectural call was implementing it and finding out. With AI, the cost is implementing a verifiable form of it quickly enough that "is this the right bet" becomes a 24-minute question instead of a 3-5 day commitment.
24 minutes is not a productivity metric. I have no pre-AI counterfactual for this project — I didn't write a Rails-to-native-binary transpiler in retirement before this; that timeline doesn't exist. What 24 minutes is, instead, is an exploration budget. Every architectural fork in the path can be probed. The ones that turn out to be dead ends cost a session; the ones that work yield commits.
I'm retired. I could be in a rocking chair reading novels with predetermined plots. Instead, I'm playing a choose-your-own-adventure where the "choose" is mine — which idea is worth chasing, which target makes sense, which architectural metaphor to borrow — and the "adventure" is built with a co-explorer who doesn't get bored or impatient.
I'm learning web technologies, compiler technologies, and ORMs simultaneously. Not in curriculum order. Not because I have to. Because it's a richer plot than the predetermined ones.
The disclosure
This isn't novel and isn't fringe. The Ruby AI newsletter on April 30 summarized Matz's RubyKaigi disclosures: "roughly a third of Ruby core committers already write 50 percent of their code with AI, and he says he writes 100 percent of his code that way." The same newsletter describes Spinel as "an ahead-of-time Ruby compiler that produces standalone native executables, built in roughly a month with Claude as co-author." Matz's own commits on matz/spinel carry the Co-Authored-By: Claude trailer. Roundhouse's commits do too.
Two human principals on opposite sides of the channel; the same assistant; transparent attribution on both sides. That's not coincidence. It's the visible shape of how Ruby's most-influential maintainer and a retired Apache developer happen to be working on adjacent projects in 2026.
A third of Ruby core writes half their code this way. Matz writes all of his this way. The phenomenon is bigger than the commit trailers; the trailers are just where it surfaces.
What this is
Some readers will flinch at the AI-as-co-author framing. The discourse is fraught. The alternative — pretending the work happens without the assistant — would be dishonest, and both projects on this channel have already declined that framing.
I'm not selling. I'm not benchmarking against pre-AI productivity. I'm reporting an experience: the choose-your-own-adventure game where every fork in the architectural path is probeable, where the cost of trying drops to where ideas worth a half-day of investment yield 24 minutes of result, where the patient co-explorer never gets bored.
For someone who could be reading novels in retirement, that's a richer plot than the novels.
Roundhouse is open source: dual-licensed MIT / Apache-2.0. Issues and discussion welcome.