[cadamsdotcom]

Every task is different. There are some “dimensions of complexity” that affect where to operate on the “vibecode/handcraft” spectrum..

First is the degree to which your target framework, language, and domain are in-distribution for the model. You’ll get far rather in python than in Verilog, for example. You’ll get further vibecoding next.js than whatever people use for web apps in Elixir.

Second is the amount of context gathering. A greenfield project has no context - every project starts from the same zero point: an empty repo or generated scaffold. Large codebases must be loaded into working memory even for humans. This is why professional software engineering depends so heavily on getting into “flow”: https://i.imgur.com/3uyRWGJ.jpg

It’s just horses for courses.

My prediction is LLMs will get there; they’ll scale to larger and larger codebases as context windows get larger, and working out-of-distribution will happen thanks to scaling inference-time compute and agentic capability to research, read code, build understanding, and store said understanding in a scratchpad dedicated to you.