What they did
Developer 박성상, who says they are not a full-time developer, spent a few days describing what they wanted in Korean while Claude Code did the planning, coding, deployment, debugging, images and short videos. The rule was no server or API bill. The site is static HTML on GitHub Pages. Login, bookmarks, a visit counter and web push run on Cloudflare Workers and D1 free tiers, and a fridge-photo ingredient tool uses a vision model on Workers AI with a daily cap. Data comes from publishers' RSS feeds where robots.txt allows, a public crypto price API and Google Trends. A Windows scheduled task on their PC collects, processes, builds and pushes to git every hour. Features include grouping one story across 15 outlets, a verification-signals score per article and analyst reports rolled up into per-stock trees.
How it went
The site reached about 1,300 pages and rebuilds every hour with no hosting or API cost; the domain and the AI subscription are paid separately. The agent got several things wrong. Story grouping first chained matches and merged unrelated stories, fixed by comparing every article to a seed. An undated, year-old article showed up as today's report, fixed by resolving dates from several places and dropping anything older than a week. Single quotes used for emphasis in Korean headlines were treated as quotations.
Worth knowing
The hourly job runs on the author's own PC, so the site only updates while that machine is on. Generated character images had odd hands, and only hand-checked ones are used. The full instructions and cost table are on the author's build log, in Korean.