This week pushed me beyond simply following steps into actually understanding how data flows through an AI automation. Replacing JavaScript nodes with LLM chains sounded straightforward initially. In practice, I learned that LLMs silently break implicit assumptions unless context, schemas, and validation are handled explicitly. I ran into multiple dead ends like rate limits, broken loops, partial logs, and at first it felt like I was doing something wrong. Over time, I realized those failures were exposing gaps in how I was thinking about guardrails and observability. By the end, I was able to replace the JS ranker with an LLM-based ranker, add post-LLM validation, and build a more robust topic-finding workflow. Biggest takeaway: building AI workflows is less about prompts and more about protecting data flow and designing for failure.