Field Notes: Thinking with Claude

The founder's own field notes from the chat/Projects layer (distinct from the Claude Code agentic layer the product was built on). Verbatim, lightly cleaned, as a starting lesson:

1. Project instructions beat custom styles for consistency. Styles are global, instructions are scoped. If you want every chat in a project to sound a certain way, put it in the instructions.
2. Knowledge files go stale in your head. An old brief left in a project's knowledge quietly poisons answers for weeks. Clean your knowledge like you clean a fridge.
3. Starting a fresh chat inside the same project is underrated. Long chats get muddy; a new chat keeps the context and drops the mess. You will not lose the project knowledge.
4. Sonnet is fine for most project work. Defaulting to the biggest model out of habit burns limits by mid-afternoon. Move routine work down a tier.
5. "Say I don't know instead of guessing" in the instructions actually cuts confident-wrong answers.
6. One project per actual project. A mega "work" project becomes a junk drawer; splitting by client or product sharpens everything.
7. Paste your own writing into the knowledge to match your voice. Telling it "write like me" does little; showing it three samples does a lot.
8. It does not remember across projects. Context does not bleed between them.