the difference between a good CLAUDE.md and a bad one:
bad → "be a senior engineer. think carefully"
good → "IMPORTANT: run type check after every change"
one is a wish
one prevents a specific mistake
your AI is only as sharp as the constraints you give it
> typed the same extra instructions 30 times this month
> "keep it under 1300 chars"
> "single image only"
> "write it in my voice not yours"
> found out you can save all of that as a /skill
> type /newsletter and Claude already knows everything
> prompting is dead. saved workflows are the meta