Using Claude Code for over two months, it has grown from a configuration file into a full-fledged operating system.

The most painful pitfall I encountered: files in the rules/ directory are fully loaded on every conversation. I inserted 17KB of rules, which directly overloaded the context window—125,996 / 125,999 tokens—Claude couldn’t generate any output. I had to reduce it to 6.6KB to restore normal operation.
This experience taught me a design principle: every byte has a cost, and loading on demand is the correct approach.
My current structure is three-layer:
(Always loaded, <200 lines, only pointers)
→ rules/ (auto-loading, behavior standards, debugging processes, capture rules)
→ docs/ (load on demand, heavy documentation, read only when needed)
On top of this, I run four mechanisms:
Hot Data Layer —
Records progress daily, automatically writes before closing the window, without waiting for you to say "save." Next time you start a conversation, Claude can continue from the breakpoint.
Task Routing — Sonnet handles daily tasks, involving funds/strategy automatically upgraded to Opus, cross-verified by outsourcing to Codex or Gemini if needed. Four-layer scheduling, each with clear trigger conditions.
Experience Backtracking — When encountering a bug, the first step is to check the memory library; debugging without doing so is a violation of the process. Corrected errors are immediately written in.
Validation — Before claiming "fixed," run tests, read outputs, and confirm success. Never say "probably okay."
After two months of running, my biggest feeling: it’s not just a configuration file you set once; it’s a living system. When you correct it, it remembers; when you hit a pitfall, it accumulates; when you close the window, it saves itself. The more you use it, the more intuitive it becomes because it evolves with you.
What does your system look like?