An adapter for prompt-backed actions that returns the raw llm response

Useful for example for llama-guard3 which returns either:

• "safe"
• "unsafe category" as a response.