there's a really high chance you are overengineering your llm tooling and it would be better to just prompt directly with as little context as possible