As massive as I find the productivity boost from AI tools for coding (> 2x for me conservatively), with current capabilities I find it's a net negative for writing prose, even technical prose.

The problem is: I like to think I'm a better writer than an LLM, but writing is hard. Every paragraph, every sentence requires a small shot of mental energy to get right. And what the LLM suggests is never bad. It's always like, "yeah, that could work." And that's the problem. It's good enough to be seductive. To make me want to skip that little bit of effort and auto-complete the sentence, auto-complete the paragraph.

But the end result when I do that is missing something. It's grammatically correct and substantively correct. It's fine. But it doesn't grab the reader and pull them through. It's text that remains text and keeps the reader at a distance.

The core problem, I guess, is the lack of a human voice. There's some kind of essential weirdness that is missing. This generally isn't a problem for code. In most cases, code that is boring and generic and anodyne and does the job it's supposed to do is good code.

It will be interesting to see how this changes as LLMs continue to progress. Is this a fundamental limitation of the technology or a minor hurdle that will be quickly overcome?

If I could write docs with AI that would genuinely pull the reader in and hold attention better than my own writing, I'd be happy to do so. It's not sentimental for me. But for now, Copilot will stay disabled for markdown files.

1 - https://github.com/plandex-ai/plandex