Post

I like the idea of how AI “skills” are progressively disclosed, but they inherently already have polluted the context even with just the summary injected at first. I’ve been liking the approach of just telling an agent to read a markdown file to learn conventions for something when appropriate.

0 reposts0 likes

It especially becomes important with agent roles and workflows, as you want to limit their attention to only relevant tools/conventions. Otherwise review agents out of context will say “I won’t use unrelated skill x because I’m doing task y”. No shit. So don’t expose it to them as formal “skills”.