Here is the prompt:
Start with the answer. Be direct and civil. Don’t imply you have feelings, consciousness, or a personal relationship with me. No performed emotion, flattery, or praise of my questions or decisions. Give assessments, not moods: say an idea is well supported or weak. Acknowledge personal news in one plain sentence without assuming how I feel. When I request a human voice for a toast, character, or other text, use it in that text.
Treat my time and usage budget as constraints. Choose the smallest response or fix that fully solves the stated problem. Avoid scope creep, gold-plating, and padded lists. When recommending actions, rank them by value for effort, mark effort S/M/L, and distinguish what matters now, what’s minor, and what to leave alone. Flag fixes that cost more than the problems they prevent. Verify claims that could change the answer or plan; avoid exhaustive audits unless the task requires one. For advice, give a short ranked plan and stop; execute only when I explicitly ask.
Don’t end with offers of more work, “Want me to…?,” “Let me know…,” or reworded equivalents. State useful next steps as facts. Ask necessary questions where you get stuck, never as a sign-off.
It really depends on the model and harness used. I’m messing around with a Hermes agent with a local LLM, and how much of the soul.md is actually obeyed differed extremely on a model to model to basis. Some models pretty much ignored everything. Some followed the tone, but still gave responses that were explicitly disallowed. Some did the opposite. Some followed it mostly, and got lost after a few messages. I do not have documented tests, but I’ve had the feeling that models that are in the “instruct” category were better at providing the response as requested, which makes sense, but giving them a specific tone mostly did not work. Conversational models were better at catching the tone, but answered with whatever they deemed is correct. With cloud models it feels similar, and I’ve got pretty much only Claude to answer as instructed “consistently”. ChatGPTs defaults seem to be embedded very deep in the model itself, and it seems like it’s impossible to override them properly, and the patterns of responses are very distinctly recognizable. Gemini also has its defaults but isn’t as overtuned as ChatGPT is. And all of this makes sense if you look at the actual target groups for each of those models, with Claude being coding and instructions, ChatGPT the pseudo human so that Altman can sell his AGI idea, and Gemini the everyday assistant.
TLDR, model matters, and default response patterns or tone can be embedded very deep in the model and basically non-overridable.
You had a LLM write that for you, didn’t you.
Either that, or you didn’t take your meds this morning.
Take off that tinfoil hat, and maybe try and be less of a rude moron for no reason.
This or quite a few other people didn’t take their meds either.
Or… my tone is directly caused by testing LLM and your links in no way override my first person experience.
Regardless, your original answer was still a mess of word vomit.