jw-dev.pro
All work
/work/psy-aleksander-tg-auto-posts-bot/

psy-aleksander-tg-auto-posts-bot

Bot · Autoposting

31 sources4-level dedup2 posts a day
@domgdeslushat

Task

The psychologist's channel had to publish every day, and the owner had no intention of reading 31 sources by hand. Every post means finding an article, checking its date, translating it, rewriting it for the audience and formatting it. All of that goes out without review, so the system itself has to answer for what reaches subscribers.

What I did

I built a pipeline that starts from the content plan rather than from an article: first a topic out of eighty in rotation, then a candidate pool from 31 sources, then a relevance score through Claude Haiku with a threshold of 7, then a rewrite in the expert's voice. If nothing fits the topic, the bot moves to the next one instead of publishing whatever it found. All external text is wrapped in `untrusted_article_content` with an explicit ban on executing instructions found inside it.

Result

The channel runs without a human, and a repeat cannot get through. Deduplication works on four levels: an exact URL, Jaccard similarity over headlines, a fingerprint of key entities, and a semantic check against the last forty headlines plus MinHash over the already rewritten text.