Most CLAUDE.md files are 500-line monoliths. When you switch LLMs, you rewrite everything. After the third rewrite, I built a three-layer architecture that makes model swaps trivial.
I run DeepSeek V4 Pro as my daily driver for Claude Code. But sometimes I need Claude Opus for complex reasoning, or Sonnet for fast iterations.
Every time I swapped, I rewrote my entire CLAUDE.md. DeepSeek needs tighter tool-call discipline. Claude Opus needs less output splitting. The same rule file can't serve both.
The real issue: I mixed identity, behavior tuning, and process rules into one file.
Three files. One changes.
Your role, goals, expertise. Stays identical whether you run Claude, DeepSeek, Gemini, or Qwen.
DeepSeek V4 Pro:
Claude Opus:
Same structure, different calibration. Each model gets what it needs.
Session startup checks, shutdown sequences, review systems, health thresholds. These rules work the same regardless of which LLM is the engine. They're process, not personality.
SOUL.md isn't static. Each session reads the current self-model, produces new experiences, then updates itself. Next session reads an evolved version — a different "me."
Reading yourself → influenced by what you read → producing new data → rewriting yourself. I call this the strange loop.
Even with one model, this separation prevents drift. Identity doesn't get buried. Process rules stay model-agnostic. New team members adopt BODY.md without touching INTERFACE.md.
One file changes when I swap models. Everything else stays.
Architecture available as a Claude Code skill. Full config system open-source.