Theo ships a Rust tsc he never read
Theo says five months and about $400k of Codex tokens got him nowhere on a Rust rewrite of the TypeScript compiler, and two weeks of Opus at about $20k finished it. He has not read a line of it.
The big disagreement
dax says agents keep undoing decisions nobody wrote down
dax says slop is mostly solved and the next problem is people accidentally undoing functionality, because nothing records that you rejected solutions one through four and chose five. He says tests do not capture that, and another document just gives you one more thing that drifts. Matt Pocock replied that this is what architecture decision records are for.
Code is the materialized view of all the decisions made
Receipts
Theo ships tsc-rs, an entire Rust TypeScript compiler he never read
Theo released tsc-rs, a drop-in Rust rewrite of the TypeScript compiler, type checker and LSP that agents built over five months. He says roughly $400k of Codex tokens got nowhere and roughly $20k of Opus finished it in two weeks, about ten weeks of usage on the $200 Claude sub.
I have not read a single line of the code.
How they actually work
Boris Cherny: no secret to prompting, just goal, effort, verification
Boris Cherny posted his actual prompts after people were surprised how plain they were, and says you should talk to Claude the way you would a coworker. His three things to communicate: what you want done, how much effort to spend, and how the model should verify it did the right thing.
There's no need to be overly scaffolded or prescriptive for most tasks -- give Claude a goal, and it will figure it out.