@trq212 How timely! I had set up my own Telegram hook, but had to set up a listener that would check every 5 minutes for messages. This seems like it would solve that!
@nanomader@AymericRoucher@petergostev In any context, assuming you are always right is not great. If the model is always responding in the affirmative, then you won't get the pushback you might need on bad or incomplete concepts.
@claudeai Is there a means to connect to a remote Claude Code session using Claude Code from a terminal on another computer rather than the web or desktop app interface?
@tom_doerr I'll check it out. I've done something similar but from a different perspective. Rather than remove slop, reframe the output through voice instructions and using output verification through gates. Trying to shoot for one-and-done frameworks for my use case.
1/ Gotta say, it's interesting working within @claudeai and deep diving some of the model's capabilities. I've been trying to fine tune a prompt that can generate a reasonable level of determinism across a number of categories with Haiku.
2/ Tuning for Haiku, while being able to increase determinism, doesn't directly translate across Sonnet and Opus. There seems to be permission differences from Haiku to Sonnet, and Haiku has a reasonable likelihood of forgetting DO NOT instructions.
@danieljpicard @QuickBooks Seems typical of @QuickBooks. Their support has been fairly atrocious, and they hardly even recognize they have a problem today. I feel they further tried to cut costs and vibe-coded their way into this mess.
Some good racing conditions for Saturdays stages @bigwhiterally. All cars starting Saturday too! It's going to be a tight finish to the end for the championships. @ Big White Mountain https://t.co/mM68fc7fdF