TwitX

Posting regularly about tech takes a lot of time just to find topics. TwitX runs a pipeline twice a day. It scrapes Reddit, Hacker News and GitHub, keeps what matches each user's topics, and asks an LLM to write drafts using that user's past posts as style examples. The drafts arrive on Telegram and the user decides what to post. It never posts on its own.

Finding topics and writing drafts every day takes too much time. The goal was help with drafting, not a bot that posts for you.

A simple relevance score cuts the scraped items down to the top 15 before anything goes to the LLM. The prompt includes the user's past posts and which drafts they liked or skipped. GitHub Actions runs the pipeline on a schedule, so no server has to run all day.

• Relevance = keyword hits × 10 + log(upvotes + 1) × 5. The top 15 go to the LLM, which keeps the prompt small.
• On Vercel the function gets killed as soon as the HTTP response is sent, which would drop Telegram replies. A custom dispatchUpdate awaits every bot handler before responding.
• parseDrafts tries to parse the full JSON array first. If the model got cut off by max_tokens, it rescues whatever complete {} objects it can find.
• It uses Groq by default, but any OpenAI-compatible provider works by setting AI_BASE_URL / AI_MODEL, with no code changes.
• A GitHub Actions cron (09:00 and 19:00 UTC) runs the pipeline, which gets around Vercel's short function timeout.

The pipeline works end to end with Telegram. Discord delivery isn't implemented yet (it's just a stub).