Definitely worth checking out the /insights command if you use Claude Code to help figure out what skills you may need. I do this every 30 days and find it genuinely helpful.
The amnesia framing is spot on, but I'd push further: skills also fix the inverse problem, which is agents being overconfident about codebases they've never actually seen. Without a skill file, an agent will happily hallucinate a plausible-sounding convention instead of admitting it doesn't know your patterns. Curious how PostHog handles staleness though - with 187 files across 28 products, what's the process when the underlying code changes but nobody updates the skill?
Good point. Staleness is definitely a trap and I don't think we have a good solution for it yet. It would be up to the teams working with those skills each day to identify staleness and fix it.
Section 2 took me the longest to learn. When a skill of mine starts collecting don'ts, it usually means I never wrote the procedure. Now I go back and write where to look and what to do instead, and most of the don'ts stop being needed.
Definitely worth checking out the /insights command if you use Claude Code to help figure out what skills you may need. I do this every 30 days and find it genuinely helpful.
Good one, just did it now!
The amnesia framing is spot on, but I'd push further: skills also fix the inverse problem, which is agents being overconfident about codebases they've never actually seen. Without a skill file, an agent will happily hallucinate a plausible-sounding convention instead of admitting it doesn't know your patterns. Curious how PostHog handles staleness though - with 187 files across 28 products, what's the process when the underlying code changes but nobody updates the skill?
Good point. Staleness is definitely a trap and I don't think we have a good solution for it yet. It would be up to the teams working with those skills each day to identify staleness and fix it.
Section 2 took me the longest to learn. When a skill of mine starts collecting don'ts, it usually means I never wrote the procedure. Now I go back and write where to look and what to do instead, and most of the don'ts stop being needed.
I think of agents as humans that forget everything everytime they work.
Yup, that's a good way to think about it.