@joshpuckett I have! I’ve been playing with some concepts around wait masking in particular for the app I’m building. I just think some of these performance tips compiled in a Reference Desk style section would be awesome :)
@mitsuhiko I actually think a large part of my human reviewer fatigue stems from having to read AI prose. They seem to love talking to each other in this way, though
@rileybrown One of my recent lifehacks has been to “interview” different models (big and small) about their experience using and understanding my app’s CLI/MCP. They usually give excellent responses and help me think outside the box as an engineer
@chipro In Codex at least, it seems to worsen every autocompaction cycle, since it seems to consider each turn after compaction as a greenfield task and lean towards being comprehensive. But just an anecdotal hypothesis
@paulg as somebody who has been a student founder then worked in university innovation, 100% agree. far too much time is spent teaching folks with little enthusiasm to build about business fundamentals they will never use
@theesabina It’s also the insanely verbose “explainer text” it adds to ui elements. “Off, turns off the feature. On, turns it on app-wide.” Yes we’re smart enough to understand what a toggle is, Claude
@ryancarson@paper I find it much more steerable than Claude Design, and its MCP comes with pre-baked “anti-slop” design instructions. It does take a fair bit to get your existing design language ported over but they’ve added design tokens support recently
@bcherny@mattpocockuk@JohnOusterhout I wonder if “engineering” vs “coding” is a harness vs model concern. It seems to me that raw models tend to react tactically to most requests, and it’s harnesses that make the relevant context for long-term engineering available to the model.