Prompt Caching

Prompt caching is the easiest 5× cost win in production. Most teams botch the key design and never see the hit rate they were promised.

The model providers do the heavy lifting — your job is to put the right content above the cache boundary and to measure whether the cache is actually hitting.

When to use this skill

• The user has a stable system prompt or tool list ≥1k tokens on a hot path.
• The user reports cost or latency problems on repeated calls.
• The user is comparing Anthropic, OpenAI, Bedrock, or Gemini for an agent.
• The user mentions "cache hit rate" but cannot quote theirs.

Decision flow