From the way you describe it, our process does not sound that different, except that this

> If AI were a thing that could reliably pluck the abstract ideas from my head and turn them into the corresponding lines of code, i.e. automate the "line by line" part, I would use it enthusiastically. It is not.

... is exactly how this often works for me.

If you don't get any value out of this at all, and have worked with SOTA tools, we must simply be working in very different problem domains.

That said I have used this workflow successfully in many different problem domains, from simple CRUD style apps to advanced data processing.

Two recent examples to make it more concrete:

1) Write a function with parameter deckName that uses AnkiConnect to return a list of dataclasses with fields (...) representing all cards in the deck.

Here, it one-shots it perfectly and saves me a lot of time sifting through crufty, incomplete docs.

2) Implement a function that does resampling with trilinear interpolation on 3d instance segmentation. Input is a jnp array and resampling factor, output is another array. Write it in Jax. Ensure that no new instance IDs are created by resampling, i.e. the trilinear weights are used for weighted voting between instance IDs on each output voxel.

This one I actually worked out on paper first, but it was my first time using Jax and I didn't know the API and many of the parallelization tricks yet. The LLM output was close, but too complex.

I worked through it line by line to verify it, and ended up learning a lot about how to parallelize things like this on the GPU.

At the end of the day it came out better than I could have done it myself because of all the tricks it has memorized and because I didn't have to waste time looking up trivial details, which causes a lot of friction for me with this type of coding.