full talk: Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic
youtu.be/vlfCjWuv4es
52
anthropic's claude code team just said the quiet part out loud:
your CLAUDE.md is probably working against you.
lydia hallie (claude code team) went through the old best practices -- "IMPORTANT", "never do this", exclamation marks everywhere -- and said the quiet rule out loud: with today's models you don't need that, and it can actually make the model go off track. more tool calls. worse results.
the new rule: your CLAUDE.md should only contain the things claude CANNOT see from your code.
stuff like stack, where to look, what folder does what -- the model already knows it.
and there's a command that does the audit for you: /doctor prompt-audit.
it checks your CLAUDE.md and your skills and tells you what the new model no longer needs
run it before your next session. takes a minute, saves hours.
3
126
Turns out well written quality content and DR actually put you into AI overview.
Who would've thought
The app is not even on Slack App marketplace yet!
Got myself a 40 points checklist to write a content, now I have ai check it for each article
32
hermes + kanban + wisprflow
i may have been too harsh on hermes. once you figure out profiles it's so easy to work in it.
but working on mobile is still impossible, if youre not on desktop youre ngmi
my Hermes agent can now check whether its own actions comply with EU regulations
asked if that was a conflict of interest
it spawned a subagent for "independent expert assessment"
great. i now have eu parliament on my mac mini
69










