The skills system is essentially claude.md but structured for human-readable reuse rather than session injection. I've been running a version of this for months.

What I found interesting is that the underlying model matters more than you'd expect for skill execution. Same skill file, different model - meaningfully different quality of autonomous decision-making. With Sonnet 4.6, I noticed the morning routine skill started making judgment calls I hadn't explicitly coded for. Not wrong calls, just more autonomous ones.

Your three-tier structure (morning, evening, weekly) is solid. One thing worth adding: exception handling. What does the agent do when the daily pattern breaks? That's where most skill systems quietly fail.

Related experiment from this week: thoughts.jock.pl/p/sonn…