Deployed an automation that search the web for content ideation for my blog/videos daily at 7am. There were couple of things that broke during this learning/building session:
1 . The model was hallucinating some of the URLs it attached and what Claude did was harvest the real citation URLs directly from the web-search results (the structured data), append them to the research digest as an explicit "real sources" list, and tell the structuring step to use only URLs from that list — drop any idea whose real URL it can't find.
2. Duplication of ideas - Claude fixed it by adding a deduplication step
3. Search agent got timed out - Claude fixed by increasing the headroom