cache-audit

To audit prompt caching configurations, run a cache audit that checks settings.json and CLAUDE.md for dynamic content placement, hook outputs, tool stability, and model switching. This generates a scored report to identify and fix issues causing unnecessary token usage and cache misses.

Frequent cache misses and high latency often stem from prompt caching anti-patterns like dynamic content injected into static system prompts, mid-session tool changes, or unstable compaction handoffs. Auditing your config detects these placement issues so you can stabilize session prefixes.

Check if hooks cause cache invalidation by verifying they emit dynamic data via additionalContext or system-reminder messages instead of directly editing system prompts. A cache audit reviews hook files and session configurations to detect this anti-pattern and prevent unnecessary cache misses.

The best way to validate MCP tool defer patterns and model switching is to run a cache audit that detects mid-session tool or model changes and estimates per-turn injected content size. This checks tool schemas and compaction patterns to ensure consistent prefixes and reduce token costs.

Prompt caching in Claude Code works with git status injections and local user settings when static system prompts load before dynamic content. A cache audit reads ~/.claude/settings.json and CLAUDE.md to verify ordering, estimate injected content sizes, and ensure session-specific data does not break the cache.

Run a prompt cache audit during onboarding, CI checks, or debugging when sessions show frequent cache misses, unexpected behavior, or high token costs. The audit reviews skill definitions and MCP configurations to ensure consistent prefixes and tool schemas across compaction and subagent handoffs.