What's this about?
Following his earlier observation that even a strong model tends to optimise within the frame you hand it, dotey tried a different prompt: "if you stepped outside the current architecture and redesigned this pipeline based on what it needs to do, what would you do differently?" It produced genuinely different ideas. He then made the model prove it — spin up a git worktree, implement the alternative, and benchmark it against the current one without sacrificing quality. The verdict: not as good as hoped, but not worthless either, and partly adoptable.
What is a git worktree, and why use one here?
A worktree is a second working directory attached to the same repository, letting you have two checkouts side by side without switching branches or disturbing anything running. He used one because other tasks were already in flight. The reason it matters is what it enables: instead of arguing with the model about whether its idea is good, you make it build both and measure. The model's proposal becomes a testable claim rather than a persuasive paragraph.
What's the community saying?
- His three conclusions, in his own order: asking the model to step outside the frame does produce useful new approaches.
- But the proposal is not necessarily sound, and needs human verification, preferably with data — don't accept it wholesale.
- And a workflow note: he moved the "explain this to me properly" conversation into a sidechat (`/btw`) to keep the main context clean, only bringing the explanation back once he understood it.