I found this interesting. But the best approach is start with LLM, then check every point yourself, and summarize with real links. The moment we are ok with LLM output just once, it won't be just once, and things get too murky.