Post

Ok, another one. I'm trying to get claude to implement a 3D calendar which matches my mental imagery for time. It's a recursive coil, like this. Claude is fine at common patterns like "implement a cone mesh". I expect it has O(0) examples of this pattern in training, though

hmm. I suspect I could get codex to do this, which means I also suspect you could get claude to do this. is your prompting cad-like? parameterization and constraints?

Hmm, what do you mean by "CAD-like"?

I'm testing having claude do it. here's my first real prompt:

result of prompt 1: I realize the first coil is the easy part, but it's good to make sure the model understands things spatially first. it's easier now to just refer to "color segments" or something

we have achieved hour coils. now... getting to seconds... that's where this becomes like a perf problem rather than a "can claude do it" problem

damn! what were the rough steps to get from your first prompt to this?

I mean really once I had the first coil it's very easy to ask for a coil around the coil. that's basically what I did. told it the ratio of turns (1 year coil turn to 365 day coil turns, 24 hour turns to 1 day turn), and asked for a coil around the coil with an offset parameter for each

the part that makes this like a multiday project in my mind is what the actual UX would be like. I'm thinking like the camera has a fixed position and the coil rotates/moves through its view, so you can do LOD stuff and cull the finer time coils away from the "focus spot"

yeah, I have been thinking ahead to that stuff, so the task has been "compute a general transform between (scale i, time t0) and (scale j, time t1)" and it is getting more confused. but it feels like that's too much for it to juggle so I am starting over trying to just get one level right

in general I think it's best to treat working code like part of a prompt. if you have some more clever idea for the actual way the code would work in the final version, it's best to get something that works right, and then point to that project in your final pass where you "convert" to your method.

in your case I'd build two versions. the version with the right recusion where you don't care how its implemented, and the version that does 1 level right with your specific implementation. then the final version is combining the two.

🔥 this is giving me some new ideas

also, the restart attempt is going much better

claude is beginning to understand. it is fragile. it gets confused by the transform logic very easily