miniblog.

LLMs are surprisingly good at reducing crash samples. I've had success with "this project crashes my static analysis tool with the following command, try to shrink the repro whilst preserving the crash. Commit each smaller version to a branch".

ASTs typically discard comments, and that's usually what you want. The only time (AFAICS) that preserving comments is useful is for writing a code formatter. Could you write a formatter in terms of a list of lexemes? A CST is a non-trivial bit of code for one use case.

Underrated perk of JSON: it's easy to programmatically make changes to a file made by a human. Very few other formats have this advantage. `cargo add` took a bunch more implementation work than `npm install`, because preserving TOML comments takes more work.