菜单
Analyze prompt cache hit rates, estimate cost savings from cached system prompts and memories, and recommend improvements to caching strategy. Use when checking cache performance, investigating high token costs, optimizing cache hit rates, or diagnosing slow cache warmup.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | cache-inspector |
| description | Analyze prompt cache hit rates, estimate cost savings from cached system prompts and memories, and recommend improvements to caching strategy. Use when checking cache performance, investigating high token costs, optimizing cache hit rates, or diagnosing slow cache warmup. |
| version | 1.0.0 |
Inspects the Claude prompt caching system to report hit rates, cost savings, and optimization opportunities. Prompt caching saves 90% on re-reads of large content (system prompts, memories, tool definitions).
| Use this skill for | Use a simpler approach for |
|---|---|
| Investigating why a session's token cost is high | A single expensive read you already understand — just cache that content, no analysis needed |
| Diagnosing a low or slow-warming cache hit rate | A fresh session that has simply not warmed yet — wait 2-3 messages before measuring |
| Deciding which memories/prompts are worth caching | Content you already know is under 1024 tokens — it can't cache; expand or merge it directly |
| Producing a cost/hit-rate report or tracking trend over sessions | A one-off "is caching on?" check — read ~/.claude/settings/prompt-caching.json |
Start simple. Reach for this skill only when cache behaviour is actually costing tokens or you need to measure/optimize it — not to confirm a single obvious cache decision.
/cache-inspector [action]
/cache-inspector # Show current status (default)
/cache-inspector status # Detailed cache status
/cache-inspector analyze # Performance analysis with trends
/cache-inspector optimize # Get actionable recommendations
/cache-inspector report # Full report (save to file)
/cache-inspector clear # Clear cache (for testing only)
Each action's full behaviour, options, and exact output shape are in references/actions.md — read it for the detail behind the summaries below.
| Action | Purpose |
|---|---|
status (default) | Current cache state: active entries, session hits/savings, health verdict. |
analyze | Deep performance analysis with hit-rate trend, cost impact, and optimization opportunities. |
optimize | Prioritized, actionable recommendations to raise the hit rate, with an estimated score delta. |
report | Full report (history, cost breakdown, inventory, TTL timeline, score); --save writes to .claude/learnings/cache-performance.md. |
clear | Clears all cached content (testing/stale-cache only). Confirms first; next session re-warms at full cost. |
For how Claude's prompt cache behaves (what's cacheable, the ≥1024-token minimum, TTL tiers and
pricing), where it is configured (~/.claude/settings/prompt-caching.json), and the target
metrics that define a healthy cache, read
references/caching-internals.md.
Low hit rate, entries expiring fast, or "no cache data available" — see references/troubleshooting.md.
Always
prompt-caching.json is configured before claiming "no cache data".Ask first
clear — clearing all cached content forces a full-cost re-warm on the next session; confirm before running.~/.claude/settings/prompt-caching.json (shared, session-wide config) or moving content into memories via /context save.Never
clear as a shortcut to "fix" a low hit rate — diagnose the cause first (it only makes warmup cost more)./context — depends-on: load memories to maximize cache hits (pre-warm before measuring here)./optimize — heavier-alternative: full optimization mode that includes cache tuning.Manage Serena memories and session context — load, save, refresh, inspect, and clear project knowledge
Initialize Claude Code optimization for a new project — detect stack, create memories, generate constitution, configure settings
Reduce token usage by selecting a planning strategy, loading Serena memories, enforcing symbol-first exploration, and reporting savings. Use when starting a non-trivial task, asking to reduce token costs, requesting efficient execution, or wanting optimized Claude Code workflows.
Search the web for latest Claude API changes, compare findings to existing documentation, and apply targeted updates to keep prompts and configs current. Use when docs are outdated, after a Claude API update, to refresh system prompts, or to validate documentation accuracy.
Close the loop on a skill/prompt/convention library: mine recurring review feedback into rules, fold them back into the generator, and gate every change through a three-tier evaluation (deterministic lint + trigger accuracy + LLM judge) that converges instead of drifting. Use when a correction keeps recurring across skills/PRs, when adding or revising skills warrants a quality gate, or when asked to "harden the skill library", "run the eval", "lint skills", or "close the loop". Composes with your existing execution, memory, and governance layers.
Preserve the OPERATIONAL THREAD across coding sessions with a repo-local, inspectable .continuity/ surface and a resume -> work -> finalize lifecycle. Not another memory store — a thin lifecycle layer over the memory you already have. Use at the START of substantial work on a repo ("resume", "where were we", "pick up") and at the END ("finalize", "hand off", "wrap up") so the next session starts from bounded, evidence-weighted state instead of re-orienting from scratch. Complements (never replaces) Serena/Svod/codebase-memory/auto-memory.