Zevi Arnovitz on Vibe Coding as a PM, Multi-Model Review, and Building with Claude Code

Zevi Arnovitz, PM at Meta, in conversation with Lenny Rachitsky. Zevi has zero technical background (music in high school, no tech army unit) and started building apps a year before recording using Cursor and Claude Code. He walks through his complete /commands workflow live, builds a feature end-to-end during the episode, and shares the multi-model peer review technique he developed for reviewing code without being able to read it.

Key ideas

The /commands workflow replaces the need to understand code. Seven phases — create issue, explore, plan, execute, review, peer review, update docs — give a non-technical builder a repeatable loop that produces production-quality output. The phases are encoded as reusable prompts in the codebase and invoked with /command-name.
Multi-model peer review is the practitioner’s solution to the code-review problem. Zevi runs three models simultaneously (Claude Code, Codex, Cursor Composer) and uses a peer review /command to have the primary agent adjudicate the others’ findings as “other dev leads.” The primary agent has more context and either explains why issues are not real or fixes them.
Prompt iteration compounds. When AI makes a mistake, Zevi asks it what in its system prompt caused the error, then updates the prompt. This produces a CLAUDE.md that gets sharper over time. The /commands are similarly iterated. The gap between mediocre and expert AI use is largely in this feedback loop.
The gradual on-ramp matters. ChatGPT project → Bolt/Lovable → Cursor light mode → dark mode + Claude Code. Rushing to Cursor without exposure to the concepts produces confusion and abandonment. Exposure therapy for code aversion.
It is the best time to be a junior. The job market for junior roles has contracted, but the ability to build a startup solo from day one has never been higher. The net is positive for curious, hardworking people willing to learn.

The seven /commands

Command	Phase	What it does
`/create-issue`	Capture	Quickly captures idea/bug into Linear via MCP; minimal questions
`/exploration-phase`	Understand	Pulls Linear ticket, reads codebase, asks clarifying questions before any code
`/create-plan`	Plan	Produces markdown plan with TLDR, critical decisions, and status-tracked tasks
`/execute-plan`	Build	Executes plan; Zevi uses Cursor Composer for speed, Claude Code for thoroughness
`/review`	Review	Claude reviews its own output; returns critical/high/medium bugs
`/peer-review`	Cross-review	Adjudicates findings from other models; primary agent defends or fixes
`/update-docs`	Document	Updates documentation so future agents have better context

Multi-model peer review

“It’s basically saying, you’re the dev lead on this project. Other team leads within the company have looked at your code and reviewed it and found these issues. Don’t take what they said at face value. You have more context than them. You need to either explain why the stuff they found are not real issues or fix them yourself.” — Zevi Arnovitz

The three-model stack:

Claude Code — primary dev lead; communicative, opinionated, collaborative
Codex (GPT) — specialist reviewer; terse, solves hard bugs, not communicative
Cursor Composer / Gemini — fast execution and design; strong at UI work

Claude occasionally refuses to yield after multiple rounds: “This has been raised for the third time. For the third time I’m telling you, this is not an issue. This is by design.”

The postmortem loop

When Claude makes a mistake, the question is not “how do I fix this code?” It is “what in your system prompt or tooling made you make this mistake?” Claude reflects on the root cause; the answer updates the CLAUDE.md or the relevant /command. Each failure is a training step for the workflow.

Advice for larger companies

Zevi distinguishes his side-project setup (full control, single developer) from enterprise contexts. For PMs at larger companies: