What they did
Yang, a software engineer, built FakeHaul as a weekend project for a friend who struggled with impulse buying: a store where you fill a cart, check out at $0.00 and get tracking and a receipt showing the money you kept. Claude Code and OpenCode in VS Code built the first version; Claude Code later did audits, UX polish and the promo video; Codex did cleanup and moved hosting from Vercel to Railway on an isolated branch from a written spec. Strong models planned and reviewed while cheaper models such as DeepSeek and GLM wrote the code, and Hermes Agent coordinated across tools. The agents shared context through files in the repo, not one chat. Offline scripts generated the catalog, reviews and FLUX product images before each deploy, and Yang reviewed every batch before it shipped.
How it went
The site holds 1,800 products from 84 fictional sellers and 15,104 generated reviews. Monthly visitors went from 828 in July to 14,375 in September after coverage by Fast Company, Digital Trends, the BBC and TODAY. AI spend over two months was about $200, and hosting fell from $40 a month on Vercel to $5 to $10 on Railway. While matching the receipt image to a redesigned screen, Claude Code found the same item priced differently in the two places and fixed both with one shared module. A batch of 200 products came back samey, so 175 were kept after rewrites. Yang still cannot tell whether the site helps with impulse buying.
Worth knowing
With autodeploy on, every push was a release, so Yang approved changes before the push and kept the DNS cutover and billing changes for himself. He found out about the traffic from a hosting bill and would set usage alerts before launch next time. Generating all AI content offline means more visitors never adds AI cost.