I reviewed my context library and found that my agent was loading too much personal context by default.
The main issue was not that the information was useless, but that everything was being loaded at startup: values, life history, business context, investment rules, API notes, and memory references. That made the default context around 9k-12k tokens before the actual task even started.
I changed my CLAUDE.md from a simple reference list into a context loading guide. Now the agent only loads:
Deeper files are loaded only when relevant:
This reduced the default startup context from around 9k-12k tokens to roughly 1.2k-2k tokens, while still keeping the full context library available on demand.
Main takeaway: the problem was not “too much context,” but missing context routing. Context should be layered, not dumped into every session.