Best AI Vibecoding Setup (for websites)
Video summary
Companion notes
Build a full website for under $10 by splitting the work across Cursor, Codex, and a cheap cloud agent — the creator's exact vibecoding stack.
Architecture & specs: GPT-5 in Cursor
For high-level design, the creator calls on GPT-5 inside Cursor. He likens it to "speaking to Albert Einstein — you call upon it when you really need to," and uses it for site architecture and catching mistakes in code. The workflow is to have GPT-5 draft a spec, then commit it as "phase two documentation" that cheaper agents can later consume. GPT-5 thinking is included in the Cursor plan, so it's "not very expensive if you just use it to give a little bit of feedback."
Frontend fixes: Codex + browser annotation
Codex is the workhorse for incremental UI work. Its browser view lets the creator highlight problems like "this image sucks, fix it" or "make this single column" and dispatch parallel agents. One trick: screenshot an ugly page, paste it in, and tell Codex to "remake it for me." He builds the Hong Kong dog-friendly restaurant directory in 3–4 hours this way. Cursor technically has a browser too, but he calls Codex "a lot more smooth" — worth the subscription if you can afford it.
Drone work: cheap model on a cloud agent
The directory needs ~400–800 restaurant entries enriched with OpenRice data, reviews, and photos. OpenRice "blocks agents from accessing these sites," so a coder is wired into a remote Claude Code instance. The model is switched to MiniMax M3 (advertised in the video's sponsorship), which he cites as "less than a $1 per million tokens." The agent loops every 10 minutes until all articles are populated.
Cost reality check
TLDR from the creator: "back in the day, it would have probably cost you like hundreds of dollars just work with engineer… nowadays with AI… you're probably at max spending $10."
Watch on YouTube
Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=Mou0KQFNMcQ
