I still stand by the commands I discovered elsewhere online and outlined here, but I wanted to add a word of caution: about six weeks ago, I had to make a longish trip by car, 2 and a half hours each way, for my day job. On the way out, I manage 45 minutes of dictation, coming to about 1500 raw words, and on the way back I managed about 30 minutes of dictation, coming to not quite a thousand raw words. (For comparison, my normal dictation sessions top out at around 20-25 minutes and 600ish raw words. )

I noticed something peculiar when I tried to feed Claude the text output from those longer dictation sessions: the LLM kept trying to shorten the output, especially on the longer segment of 1500 raw words. Telling it “you are summarizing too much” seemed to help with that, but it degraded the quality of the cleanup, mostly on basic grammar corrections (wrong prepositions, words in the wrong place, things like that). The LLM continued to put in quotes and commas and so forth correctly, and it was worth it for that, but I had to do more manual cleanup than usual, which probably didn’t matter much because there was a certain amount of expanding and reworking I needed to do anyway. (After deleting repetitive bits that Claude didn’t ignore, and tweaking and expanding here and there, the equivalent part of the Hunter Healer King 3 draft stands at around 3000 words).

Anyway, just a reminder that LLMs need to be used in an intelligent way with an awareness of their limitations.