"a spec that is sufficiently detailed to generate code with a reliable degree of quality is roughly the same length and detail as the code itself" ^^ 100% don't believe in that. You should care about the code at the level of abstraction @dexhorthy is describing here. But no way in hell you can't compress it way smaller than the code itself would be (40:1 based on my measurements for a 30k LOC codebase). An LLM holds the priors for pretty much every single convention there is. Along with the cultures from Linus Torvalds to corporate Java. Two things - Understanding the model's priors for your choice of language/framework(s). - You holding those same priors. First screenshot is benchmark results from a few days ago. Second is what the 40:1 compression looks like. YMMW with typescript or python slop.