This was a really great writeup, Spike. Validates a lot of my own experiences and concerns. Nailing the balance between aggregating system prompt guidelines and keeping context available for real work is a bit of an art. I have been working on incorporating automated cycles of system prompt refinements and refactors in my workflow to help keep the kernel manageable. Also review and critic steps with fresh sessions is definitely win
I agree that a lot of this, at this point, comes down to more of an art than a science, and arguably that has often been the case (see first lesson in the MIT LISP course by Abelson 30yrs ago ;) ) .
100% on the automated cycles for refinements and refactors, and that's really what I'm moving toward with not just hooks, but actually separate agents. I think that's kind of my next step in evolution: to move quite a few things out of even skills and into agents. I've been sort of resisting this process because I didn't want to introduce more software than I had to, which I would then monitor and maintain and all of that, but at this point it seems to be an inevitable step. The big thing is that once some of those skills and prompts become agents, built on top of the Claude Code SDK to keep standard logging and hooks etc, I can actually turn them into a much more deterministic processes.
I already have a nightly for learnings consolidation and past-sessions reviews, but not enabled them in cron, just kicking them off manually at the end of my day.
This was a really great writeup, Spike. Validates a lot of my own experiences and concerns. Nailing the balance between aggregating system prompt guidelines and keeping context available for real work is a bit of an art. I have been working on incorporating automated cycles of system prompt refinements and refactors in my workflow to help keep the kernel manageable. Also review and critic steps with fresh sessions is definitely win
I agree that a lot of this, at this point, comes down to more of an art than a science, and arguably that has often been the case (see first lesson in the MIT LISP course by Abelson 30yrs ago ;) ) .
100% on the automated cycles for refinements and refactors, and that's really what I'm moving toward with not just hooks, but actually separate agents. I think that's kind of my next step in evolution: to move quite a few things out of even skills and into agents. I've been sort of resisting this process because I didn't want to introduce more software than I had to, which I would then monitor and maintain and all of that, but at this point it seems to be an inevitable step. The big thing is that once some of those skills and prompts become agents, built on top of the Claude Code SDK to keep standard logging and hooks etc, I can actually turn them into a much more deterministic processes.
I already have a nightly for learnings consolidation and past-sessions reviews, but not enabled them in cron, just kicking them off manually at the end of my day.