Kun, I need to try this mate. Would you suggest I go all in on Grok Bot? I was debating between Grok and Dots. Haven't tried either yet. I'm centered on Claude most of the time for my max sub thus far. Cheers mate.
Hi Kun thanks for sharing your experience.Learning a lot from it.Iam a software architect with 20 years of experience.What are the best resources to learn deeply about AI models to enhance my architecture skills.Thanks again for sharing your knowledge.
The bottleneck you're describing is the thing I've been circling from the other end.
I've been running one very long thread on a legal matter for months, and the same work in a fresh thread with identical documents behaves completely differently. It flagged a sixty-cent discrepancy as needing correction in a two-hundred-thousand-dollar pool. Same capability, no calibration.
What accumulated in the old thread wasn't rules. It was a sense of what matters, built from hundreds of small corrections, and it's mostly an absence. A set of things that stopped being attended to. Which is why summarisation strips it: it doesn't look like information.
So I think we're solving the same problem from opposite directions. You're extracting what can be written down. I've been finding that what can't be written down accumulates anyway, in place, if the thread lives long enough. Yours transfers and mine doesn't.
Curious whether you've tried asking your own distillation to explain why it made a call, and whether the answer is one you'd have given.
Kun, I need to try this mate. Would you suggest I go all in on Grok Bot? I was debating between Grok and Dots. Haven't tried either yet. I'm centered on Claude most of the time for my max sub thus far. Cheers mate.
Distilling yourself into a skill is the right cut. The failure I keep seeing is the handoff when that skill has to run past one session.
This distilled discussion made me LOL :-) Must have.
Hi Kun thanks for sharing your experience.Learning a lot from it.Iam a software architect with 20 years of experience.What are the best resources to learn deeply about AI models to enhance my architecture skills.Thanks again for sharing your knowledge.
Is the distill prompt available outside of grok bot?
The bottleneck you're describing is the thing I've been circling from the other end.
I've been running one very long thread on a legal matter for months, and the same work in a fresh thread with identical documents behaves completely differently. It flagged a sixty-cent discrepancy as needing correction in a two-hundred-thousand-dollar pool. Same capability, no calibration.
What accumulated in the old thread wasn't rules. It was a sense of what matters, built from hundreds of small corrections, and it's mostly an absence. A set of things that stopped being attended to. Which is why summarisation strips it: it doesn't look like information.
So I think we're solving the same problem from opposite directions. You're extracting what can be written down. I've been finding that what can't be written down accumulates anyway, in place, if the thread lives long enough. Yours transfers and mine doesn't.
Curious whether you've tried asking your own distillation to explain why it made a call, and whether the answer is one you'd have given.
I cloned the repo and ran the npx skills add, ran my codex and it did not recognize /kun on the prompt.