Don't bloat your LLM's context. Let it iteratively explore the parts of your flame graphs that actually matter.