Like many nerds, I feel so embarrassed whenever I see that I’ve misspelled something. But some years back when I started producing text in huge amounts for things like the Eclipse Comics read-through I realised that my options were: 1) Remain neurotic about this stuff and produce one blog post per week, or 2) Start caring less and produce a lot of blog posts.
I think I made the correct decision — I just don’t have time to proff-read all this stuff, so typos will abound. It’s just not feasible to put as much work into that stuff as you’d do for a properly published article. The reader gets what they pay for.
But now there’s a way to have your delicious cake and eat it too — just ask an LLM for help.
NOOOO! *audience runs in slow motion towards camera*
I know, I know. Whenever I read somebody in deep AI psychosis who thinks they’re being AI sceptical saying “well, of course you shouldn’t let the LLM write your articles. Instead I let it do the research, and then I make it provide an outline, and then I write some text based on that, and then it does formats and restates what I wrote. IT”S A WHOLE DIFFERENT THING! I WROTE IT!!!!”
No. You should never let an LLM do those things: Not the research, not the outline, not the formatting, not anything. Reading the output of LLMs is abhorrent to people. You should never subject anybody to an LLM generated text, even if you’ve done “quality control”.
LLM-generated text gives people the ick. It’s better to have an awkwardly phrased article you’ve written yourself than subjecting somebody to a Claude text. It’s not just the ick — it’s the horror.
However, you can use LLMs for spell- and grammer check:
I’ve whipped up a li’l minor mode for Emacs (and put it on Microsoft Github). The default prompt is very clear that I want the LLM to do nothing but copy edits — no structural changes, and no other helpful suggestions. Just fix grammer and tippos.
And then you can interactively go through the fixes and toggle back to the original ones if you disagree.
Yes, yes, sure — Grammarly, etc — stuff like this already exists. But I just don’t like using them. I use the built-in Emacs spellcheck, but it doesn’t catch everything. Having an LLM give the text a quick once-over for obvious mistakes seems nice.
While testing this minor mode now, I see that LLMs haven’t changed as much as I thought — they’ve gotten so much better on programming the last half year, it’s incredible. For stuff like this, though, it’s still pretty unreliable. Any two runs won’t necessarily fix up the same thing, or the same thing the same way. And getting it (I’m using Claude Sonnet 5) to output it in the requested form is not reliable at all — I’ve been reformulating the prompt, trying to get it to output it in the format I the mode needs, and… I think I’m at 95% now?
So, like all with all non-programming tasks, LLMs are still pretty much toys. But useful ones!