> Anyone who has used AI tools has seen cases where the first 80% of a project came together like a bolt of lightning, but the last 20% is next to impossible for the AI to accomplish

Offtopic: doing the inverse works better.

If you rely on AI to write 80% of your code, you end up with code you don't understand, so the last 20% is really difficult. You can't even guide the AI correctly at that point.

I actually do the opposite. I write 80% of my code, have an LLM read it, and it helps me iterate faster on the last 20% (which are mostly tedious things like unit tests, corner cases, optimizations, etc). This works surprisingly well because now the AI has a structure to start from, and it's only doing incremental code completion. As a human, I'm also not that great at detecting corner cases -- this is where an LLM really helps because it has seen patterns I've never seen.

In fact, I use coding agents to combine multiple sources of information (not just code) like design docs, LaTeX, etc. and I ask the LLM to detect if the code actually implements the math in the LaTeX correctly. Claude actually does a pretty decent job at that.