I read somewhere that Boris (created Claude Code) recommends deleting skills, and hooks every so often and then observe how the LLM performs without them.
Maybe better to periodically prune: tweak some skills, shorten some, delete some.
Skills are for me special knowledges that an agent in general doesn’t know in details: using a specific framework, create reports, new programming languages etc.
So having them not only enable the AI to work more effectively but also more efficiently and cheaper (why digging through the ten thousands lines of help again if it was already distilled?).
I don’t use SKILLs for quirks and gotcha. They go into AGENTS.md because it’s always loaded and therefore much more reliable (there is a study about it). I consider SKILLs load-on-demand even though the agents can discover it and decide to load by itself.
I have CC do a review / audit of my entire CC stack and architecture every major model / harness release for recommended changes. What goes into my CLAUDE.md files has been shrinking over time. There is a skill review tool as well which helps determine how effective they are.
Maybe better to periodically prune: tweak some skills, shorten some, delete some.