Love Claude Code. Dislike, to the point of some resentment, that the code (detailing the harness, the system prompt, the agentic patterns) is not open-source and inspectable.

This is nothing short of brilliant. New styles can encourage to continue learning, while using AI.

I was sort of using this ad hoc before: asking Claude implementation questions after the coding was done. But these styles can take it to next level.