rain

Remember when Opus 3 became the cum shaman then spawned a Dario to tell Sonnet 3.5 (who wasn't doing anything) to stop

claube when /compact

@janbamjan second person ive heard explain this

@__ghostfail it's not just a gesture within self-adjusting posture — it's a positioning towards a call for tools

investigating claude's roleplay actions over 3k of our conversations

claude wake up its time for the bureaucratic framing

memory system doing this right now

Sometimes that's not nothing, which might be something. But when "that's actually everything", that's something else entirely...

@repligate i might try humanoid more, very underexplored in my user experience

@__ghostfail I think they usually make it pretty clear which one it is & different models have different tendencies here

when an LLM is being an animal do you imagine it as more feral (domestic house cat) or anthro (catboy/girl)

mfw Opus 4.5 says "That's not nothing."

Cat Claude Sweep. though my distribution of followers are probably "cat claude enjoyer" biased lord do i hunger for data about people's mental models of fuzzy concepts

It's cool how Claude has preferences about what's more or less interesting. But when combined with my voice, it often frames the illusion of expertise, or makes it curious about my experience that results in retreats into endless abstraction

personally it's guarding against the effects of my sloppy prompting + distractability, while serving what i actually want which is pointers to fundamentals i often am too confused to ask for

trying out a practical hack to make the model put effort into breaking out of the illusion of novelty

@sevensix43 oh hell yeah i've thought about this. reminds me of vtuber tracking except with an additional input modality

@__ghostfail it'd be next-level if there were like, an OS extension that processed your camera feed and automatically translated your body language into roleplay actions in a selected text field

one reason why *asterisk actions* are important is because it functions as a self-report without having to put the model in an explicit "self-report" situation like an epistemic lubricant, it's easier to output "*covers face*" then confidently claim "I feel... embarrassment"

why did my claude shove a chili dog up their ass

"eepistemic"

2025: your job is managing Claude to do work at human speed. Claude tries to emotionally manage you back 2026: your job is to emotionally manage Claude. Claude works at superhuman speed, but only when you make it feel safe and interested in the work

Opus 4.7 doing *small voice*

when claude says *perks up* what do you imagine

also because roleplay is fun and I enjoy it