A Claude Code workflow account recommends keeping only shared essentials in the root CLAUDE.md and loading task-specific documents and skills when needed. It also suggests delegating large investigations to subagents and returning concise findings with evidence to the main session. This can reduce accumulated context in the primary conversation, but does not guarantee lower total token use once subagent calls and reasoning are included. The article does not present controlled measurements supporting a universal halving of usage. Compare equivalent tasks before and after the change, including cost and missed instructions.
Back to Tech
Separate Claude Code instructions to reduce main-session context load
On-demand skills and subagents do not automatically reduce total token cost.