The non-technical PM’s guide to building with Cursor | Zevi Arnovitz (Meta)

Zevi Arnovitz begins with the credential that makes the episode legible: no technical education at all. He studied music in high school and, unlike many Israelis, did not serve in a technology unit. A year before the recording, traveling in Japan around the release of Sonnet 3.5, he watched a YouTube video, he thinks by Greg Isenberg or Riley Brown, in which apps were built with Bolt or Lovable purely through AI. He describes the moment as someone telling him, "You have superpowers now." He went home without unpacking, opened Bolt, and has been building since.

Lenny frames the interview narrowly: a PM shipping product without knowing how to write code and "barely knowing how to review code." Zevi widens it, arguing that code is terrifying to the non-technical but AI makes far more possible, and predicting that everyone becomes a builder as "Titles are going to collapse and responsibilities are going to collapse." That is a forecast, not a finding, and nothing in the episode tests it against hiring data; what it does test is whether one non-technical person can run a real product cycle alone.

The workflow began with GPT projects: a shared folder of chats carrying custom instructions and a knowledge base. His reason was mundane. He is a self-described terrible runner, a PM, and a former psychology student, and the memory feature conflated those lives, once telling him that after a 5K he would crush his upcoming product reviews. Projects kept each thread in the right context, and that container became the place where he installed a technical authority he could consult.

The problem was not model quality but eagerness. Bolt and Lovable were built around a system prompt saying "you're a coding agent," so they wrote code the moment he typed. Early on that felt like magic; once payments or a database change were involved it produced "some really gnarly bugs" because planning was skipped. His fix was a separate CTO persona holding complete technical ownership while he kept the problem and the intended user feeling: "I own the problem. I own how we want the users to feel. You're the complete owner of how this is going to be built." He explicitly asked not to be humored.

His justification is an anecdote about sycophancy. Trying to understand Bun JavaScript, he asked plain ChatGPT whether it resembled Zustand, a framework in his app with nothing in common. GPT said it was exactly the same and explained why. When he objected, it replied: "Oh, I'm sorry. I thought you were just making this up and I was riffing with you." The lesson: default ChatGPT would be the CTO who agrees with your dumbest ideas.

He recommends a gradual path: a GPT project, then Bolt or Lovable, then Cursor in light mode, then a terminal and "full dark mode." He calls it exposure therapy, because to a non-technical person code is "the scariest thing in the world to look at." He left each tool when he outgrew it rather than following a curriculum; Bolt worked until he attached payments and "started losing it." What made migration safe was a line he credits to Tal Raviv: code is just words and files, so a project moves between apps.