# Prompt caching: find the unstable prefix

Separate shared instructions from request-specific content. Compare the actual serialized prefix across repeated requests, then use provider-reported cache metrics to evaluate reuse. Prefix stability alone does not establish eligibility or a cache hit.

## Worked example

Illustrative layout: shared policy and tool definitions come first; the changing customer question follows. If a timestamp or random request label appears before the shared material, compare the serialized requests before assuming the same prefix is being reused.

## Copyable prompt

```text
Audit these three serialized prompts: [PROMPTS]. Identify the longest identical prefix and mark every changing field. Propose a layout with stable instructions first and variable request data later. Do not remove information needed for correctness. List the provider-specific cache eligibility, breakpoint and expiry rules that still need verification. Create a measurement table for cache reads, writes, total tokens, cost and latency.
```

## Checklist

- Compare serialized requests, not just templates.
- Keep necessary dynamic context accurate.
- Verify eligibility in current provider documentation.
- Measure hits and total cost on repeated requests.

## Further reading

[OpenAI: prompt caching](https://developers.openai.com/api/docs/guides/prompt-caching)