seed-prompt-cache-anchor

Maximize Anthropic prompt caching by restructuring prompts to move stable instructions to the top and marking the cache breakpoint. This ensures only per-turn-changing content sits after the breakpoint, increasing cache hits for long-running agent workflows.

A stable prefix is the per-turn-invariant instruction block in your prompt. Keeping it above the cache breakpoint allows Anthropic models to reuse cached tokens, minimizing per-call costs in long-running agent workflows.

Verify cache hit rates by auditing the cache_creation and cache_read metrics returned by the Anthropic API. This confirms your stable prefix placement and breakpoint demarcation are successfully generating expected cache hits.

Yes, prompt caching works by isolating stable content from per-turn-changing content. You must remove variable junk from above the breakpoint so only the stable prefix is cached, while dynamic content sits below it.

The best way to reduce Anthropic API costs is to restructure prompts by placing stable instructions at the top, marking a clear cache breakpoint, and verifying cache usage metrics to prevent per-turn-changing junk from invalidating the cached prefix.

Your prompt cache misses because adding new instructions above the cache breakpoint alters the stable prefix. You must keep per-turn-changing junk below the breakpoint to maintain cache hits and minimize per-call costs.