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

A non-technical PM ships real products with AI agents

Vibe coding has a method, and a Meta product manager with no engineering background has one that ships. Zevi Arnovitz runs Claude Code inside Cursor through a loop of planning, building and adversarial review. Agents type while the human supplies judgment about what to build and what to fix.

Takeaways

• A non-technical PM can ship paying products. Zevi Arnovitz, a PM at Meta with no technical background, built StudyMate, a quiz app that turns uploaded materials into tests, using Claude Code inside Cursor. Localizing it from Hebrew to English took two days.
• Planning with the agent separates apps from demos. His commands force a conversation before code exists. The exploration phase feels like talking to his engineering manager, and that back and forth is the difference between vibe coding and building serious apps.
• Adversarial review beats one model's self-review. Claude, Codex and Cursor each review a feature, then a peer review command has a dev-lead agent argue their findings until nothing is left. He calls Claude the perfect CTO and Codex a hoodie-wearing genius.
• Fix the root cause, not just the prompt. After a mistake he runs a postmortem and changes documentation or tooling rather than only editing a command, so the same error cannot recur. He keeps updating docs as the code base grows.
• Opinionated tools trade control for speed. Bolt and Lovable removed hard decisions but also took control, so he moved to Cursor, where he decides how the app is structured. His path ran from a GPT project to Bolt, then Cursor in light mode.
• AI-native code bases need technical owners. Making a code base AI native means plain text and markdown files explaining how agents should work in each area, and he says technical people should do it. PMs should stick to contained UI projects and send pull requests.
• AI does not weaken product managers. He strongly disagrees that AI atrophies PM skills, since the job is getting the right solution to users quickly and you own your outputs. Curious, hardworking, kind communicators can now give companies more value than some veterans.

The command sequence behind a StudyMate feature

1. He describes the feature in plain language, down to 30% fill-in-the-blank questions, six answers across two blanks and drag and drop.
2. /create issue writes the ticket, STU88, to Linear through MCP.
3. /exploration phase asks the questions an engineer would raise before writing code.
4. /create plan produces a markdown plan, and Cursor's fast Composer model executes it.
5. /update docs closes the loop so the next session starts from the current state.