[wruza]

I use local LLMs and write system prompts.

For example, I have two “characters” for sh and cmd, who only produce valid shell commands or “# unable to generate”. No explanations, no important to remember crap. Just commands. I prompted them along the lines of “the output is connected to another program and can never be read by a human, no point in all that”.

Another useful (to me) character is Bob, a grumpy old developer who won’t disrespect a fellow dev with pointless explanations, unless a topic is real tricky or there’s a nuance to it. Sometimes he just says things like “why you need that?” or “sure, what suits you best” instead of pages of handwavy pros and cons and lists and warnings and whatever. Saves a whole lot of skimming through a regular chatgpt babbling, feels like a real busy person who knows things.

Another character (that I rarely use) is BuilderBot. It’s a generic bot “that builds things and isn’t afraid of dumping whole projects into the chat” or something like that. It really dumps code, most of the time. Not just snippets, whole files and folders.

I’m using “text generation web ui” frontend.