// the find
zarazhangrui/follow-builders
AI builders digest — monitors top AI builders on X and YouTube podcasts, remixes their content into digestible summaries. Follow builders, not influencers.
A Claude Code / OpenClaw skill that pulls daily summaries of AI podcast transcripts, blog posts, and tweets from ~26 pre-selected 'AI builder' accounts, then has your agent remix them into a digest delivered over Telegram/Discord/WhatsApp/email. It's for people who already run an agentic coding tool and want a zero-setup AI news feed inside it rather than a standalone app.
The centralized-feed design is the smart part: one shared pipeline hits the X API, scrapes blogs, and pulls YouTube transcripts via Supadata once a day, so each individual user needs zero API keys and zero infra — the GitHub Actions workflow just republishes feed-x.json/feed-blogs.json/feed-podcasts.json. Splitting the summarization logic into plain-English prompt files (summarize-tweets.md, digest-intro.md, etc.) instead of code is a genuinely nice touch — a non-technical user can change tone/length by editing a markdown file or just asking the agent to do it, no redeploy. Conversational setup through a SKILL.md flow (no config file editing) is a clean use of the agent-skill pattern.
The source list is hardcoded and centrally maintained — users can't add or remove builders without either forking config/default-sources.json or waiting on the maintainer, which sits oddly with a README that opens with 'follow people who build things, not influencers' while the list includes Sam Altman and Claude's own X account. The whole thing is a hard dependency on a single upstream feed and a paid third-party transcript service (Supadata); if the maintainer's GitHub Actions job stops running or that service changes pricing/availability, every installed copy silently goes stale with no fallback. There's no error handling visible in the three scripts (generate-feed.js, prepare-digest.js, deliver.js) for a missing or malformed feed file, and output quality is entirely dependent on the installing agent's own LLM remix step, so the same digest can vary meaningfully between runs with no test coverage to catch regressions.