salience-splitter

To split a bloated CLAUDE.md file, you apply governance-semantic salience tiering to separate behavioral rules into a lean always-loaded layer, while moving execution-specific details and rarely used reference content into an on-demand layer connected via imperative pointers.

Context window optimization for Claude Code memory indexes involves splitting bloated always-loaded assets into lean core rules and on-demand detail references, reducing wasted token space while ensuring consumer agents retain cold-start compatibility and full functionality.

You reduce context waste in team project setups by applying a governance-semantic split criterion that keeps core decision logic in an always-loaded layer and shifts edge case references to an on-demand layer with built-in verification checks for pointer integrity.

Splitting SKILL.md files does not break consumer agent cold-start compatibility when mandatory imperative pointers are implemented between the lean core and on-demand layers, supported by built-in verification checks to ensure split integrity.

You should use salience tiering instead of arbitrary length splitting when your context assets contain accumulated execution-specific details, ensuring behavioral rules and core decision logic stay prioritized in the always-loaded layer rather than being truncated by character limits.

You can use this context splitting approach for shared knowledge bases containing hundreds of lines of setup scripts, applying tailored split criteria for each asset type to cut context waste by up to 80% while preserving all functionality through on-demand detail references.