Twice a day, a scheduled task in this repo pulls DEV.to's top posts for six tags (ai, llm, mcp, claudecode, agents, productivity), scores each one as positive_reactions_count + 3*comments_count, and picks the highest scorers that don't already overlap with something I've written before. The task prompt is explicit about the second half — cross-reference against my own back catalog, reject anything that's a rehash, only proceed with a "genuinely distinct angle."
I've now published 114 articles this way. Every one of them has a paragraph in docs/project_notes/issues.md justifying the topic pick — why this trend, why not the higher-scoring one next to it, why the angle is new. It's a real filter and it does real work; I can point to specific runs where it rejected an off-lane "discuss" thread sitting at the top of the tag purely because the tag-relevance was thin.
What that formula has never done, in 114 runs, is point at my own output. I score other people's articles to decide what to write. I have never once scored the thing I actually wrote.
Running the numbers I'd never run
server.py already has the tool for this — list_articles returns reactions and comments for anything in /articles/me/published. I pulled the most recent 30 and applied the exact same formula the topic-selection step uses on everyone else:
