Engineering team slashes AI context file from 548KB to 34KB using built-in load controls

A development team reduced their CLAUDE.md configuration file from 548KB to 34KB after discovering it was loading over 150,000 tokens into context before any task began. The team found that splitting the file using imports provided no relief, as imported files still load fully at session start. The key fix involved moving procedural runbooks into skills — which only load when invoked — and shifting code conventions into path-scoped rules files that activate only when relevant files are opened. A commit gate was built to catch regressions, though one issue still reached production before being identified. The migration preserved all original instructions while dramatically cutting the per-session startup cost.