dmitryprg-ai Cache Analysis
Measure and optimize prompt cache performance using claude-cache-analyzer. Cache efficiency directly impacts cost — a 95% hit rate saves ~80% vs no caching.
Prerequisites
pip install claude-cache-analyzer
Quick Analysis
# All recent sessions
claude-cache-analyzer "C:\Users\Divan\.claude" --top 10
# Specific project
claude-cache-analyzer "C:\Users\Divan\.claude\projects\PROJECT_DIR" --top 5
# Grouped by project
claude-cache-analyzer "C:\Users\Divan\.claude" -g
# Export to JSON for programmatic analysis
claude-cache-analyzer "C:\Users\Divan\.claude" --export-json metrics.json
# Detailed single session
claude-cache-analyzer -s SESSION_ID
Key Metrics
| Metric | What It Means | Good Value |
|---|---|---|
| Cache hit rate | % of tokens served from cache | > 80% |
| Efficiency score | Overall cache utilization [0..1] | > 0.70 |
| Net savings | Money saved minus cache write overhead | Positive |
| Savings % | Cost reduction vs no-cache baseline | > 70% |
Performance Grades
| Grade | Efficiency Score | Action |
|---|---|---|
| A | >= 0.70 | Excellent — no changes needed |
| B | >= 0.50 | Good — minor optimization possible |
| C | >= 0.30 | Fair — review instruction structure |
| D | >= 0.10 | Poor — rules/skills may be too dynamic |
| F | < 0.10 | Critical — architecture needs rework |
Optimization Tips
If cache efficiency is low, the cause is usually rules or context that change too frequently between turns:
- Move volatile content to references/ — only loaded when needed, not cached in every turn
- Keep SKILL.md stable — frequent edits invalidate cache across all sessions
- Use path-scoped rules — rules loaded for specific files don't pollute the cache for unrelated work
- Prefer agent-decided rules over always-loaded — reduces base context size
Why This Matters
Cache efficiency = cost efficiency. On this project, 95% hit rate saves ~$119 per session. A drop to 50% would triple the cost. Monitoring cache performance catches regressions from rule/skill changes before they become expensive.