I’m testing prompt caching with Claude for a workflow that reuses a long static system prompt, then adds different project notes for each request. The shared instructions barely change, but the project context does.

For people using Claude this way, does the cache still provide a noticeable benefit? I’m trying to understand how to separate the static system prompt from dynamic project context without making the prompt harder to maintain or sending stale information.