>>54706212>half the advanced guide is about saving tokens, genuinely recommending shit like W++ instead of showing off ways to write plaintext without redundancies>the other half is retarded placebo bullshit like brackets, bizarre formatting shit, and directly prompting the AI from within the wrapped userprompt (lmao)actually a trash guide, honestly.
LLMs are not magic boxes that parse everything the same way, some formats are better than others and plaintext is the most effective way to get the bot to act how you want. saving, say, 200 tokens of context (which is only ~5% on the tiny models aisekai uses, btw) at a massive consistency loss is a terrible idea.
putting shit in brackets and saying things like "{{Char}} will use evocative and visceral language to describe blah blah blah" does literally fucking nothing if you don't have system prompt access, WHICH YOU FUCKING DON'T WITH AISEKAI. the only way to get "evocative and visceral language" is to WRITE SOME YOURSELF IN THE GODDAMN GREETING.
and the idea of using AI to write a bot for you is also incredibly fucking stupid, but this post is long enough as-is so i won't go into the details of why AI is bad at instructing AI.
in summation, this absolute dumpster fire of a guide is exactly what i'd expect from a place like reddit. it's actively harmful to use this, and your bots will genuinely be worse off for it.
