Ideally I would build a bunch of atomic skills that combine well and claude just uses them naturally when the situation arises.

I've been building transaction categorisation tools and the parallel is striking. You can have perfect pattern logic, but the model still needs to correctly identify "this situation matches pattern X" first. Sometimes it just... doesn't, even when it clearly should.

The router skill idea is interesting. Feels like pushing the meta-matching problem up a level though - now you need the router to activate reliably. Turtles all the way down.

E.g. you could have:

- a set of skills to use design patterns of a library

- a skill to add to this skill-set -- either when prompted by user or autonomously via a stop-hook

E.g. I set up this combination for design patterns for the Langroid[1] LLM-Agent framework:

https://github.com/pchalasani/claude-code-tools/tree/main/pl...

[1] https://github.com/langroid/langroid

I was really curious to read it given the comments + full disclosure my co-founder recently wrote about a similar topic ("To Tool or Not to Tool") https://blog.codeyam.com/p/to-tool-or-not-to-tool .

I wanted to see how this is similar or different with the focus on Claude Code + Skills in a more literal sense vs. tools in a more abstract sense.