BuildFeed is a feed for the signal inside tech YouTube. A pipeline watches 110 hand-picked channels, filters each day’s uploads through four quality gates, ranks the survivors, and hands Claude Haiku the top ones to summarize. The channel list stays human-picked.
Scope at a glance
Started November 2025 with a handful of RSS feeds and a thesis: tech YouTube has a discovery problem, not a supply problem. Six months later: 110+ channels across AI, VC, startup, product, design, and tools. 300–500 raw uploads filter down to 80 curated videos per cycle. Refresh: twice a day. No human in the loop.
By the numbers
110+
+
Channels
80
Curated
2×/day
×/day
Refresh
1,039
,
Static pages
6
API units
$0.003
$
Per summary
<100ms
<ms
CDN load
2/wk
/wk
Auto-added
Featured — the top twelve, re-ranked twice a day with a seven-day rotation penalty so nothing gets stuck at #1.
Systems, not screens
Discovery runs on RSS, not the YouTube search API — channel feeds are free and unmetered, so the pipeline can watch 110 sources without touching quota. A twice-daily GitHub Action fetches the day’s uploads, runs them through a four-stage quality filter (viability gate, 0–100 engagement score, 35-point threshold, diversity cap), and scores the survivors for the featured pass. Old approach: `search.list`, 8,370 API units per build. New approach: six.
The top 12 get the expensive treatment. Claude Haiku pulls the transcript, writes a two-sentence TL;DR, tags topics, and extracts key takeaways with timestamps — about three-tenths of a cent per video. A seven-day rotation history penalizes anything that recently held the top spot, so the featured slot keeps turning over. The cache commits back to the repo on schedule; Vercel serves it static. A companion script pre-renders 1,039 HTML pages for crawlers the SPA alone couldn’t reach.
Video detail — TL;DR, topic tags, and timestamped takeaways, all generated from the transcript at $0.003 a pop.
