I spent months on the wrong half. Better prompts, retries, error handling, a cleaner graph. The pipeline got solid and the output was still something nobody acted on. What changed it was the input. I stopped feeding it everything I could scrape and started feeding it only what is actually moving in the niche right now, with the reason it worked next to it. That part runs on an AI app now, but by hand it is the same thing: ten pieces in a doc, and one line each on what they promise. Three things I would check before building the next one: - Write in one sentence the decision this thing is supposed to make easier. If you cannot write it, you are building a demo. - Feed it less. Ten good examples with the reason they worked beat a thousand scraped rows. - Check the input by hand once a week for a month. A broken input never throws an error, it just produces confident nonsense. On my side this feeds content, and last month it put me past a million views while recording less than before. If anyone is building something similar, ask here and I will tell you where I went wrong first.