In recent LLM model experiments I've been thinking a lot about how to get codegen into the workflow. For a lot of tasks, conventional (javascript) code would be faster, cheaper and less prone to error than an LLM call. I'd like to have the LLM generate that code once -- then the game becomes how to encapsulate, verify and sometimes modify the code it generates.

Some inspiration:

Future of Coding: Pygmalion

The Pygmalion podcast episode of Future of Coding. This keeps turning over in my head. Particularly the 'code by example' bits, and the partial execution. Feels to me like there's some loop for an LLM to try generating a call, and then if the result doesn't parse or pass some verification process, it opens up a teach-by-example workflow.

Parameter tweaking

Silicon Jungle exploring parsing generated code for parameters. This has ties to parametric UI like dat.gui. I also see people generating things like this with Claude artifacts. Maybe parametric tweakability is another piece of the puzzle.

Reactive flows

Orion Reed on stitching elements reactively together. This is a slightly different path, I think, but lots of great patterns.

Starting point

I need to pick some simple-but-not-too-simple task, and start working through the workflow.