How a Meta PM ships products without ever writing code | Zevi Arnovitz
Zevi Arnovitz, a non-technical Product Manager at Meta, shares his complete workflow for building and shipping sophisticated applications using AI tools like Cursor. He details a structured, multi-step process that leverages different AI models for specific tasks, including a novel "peer review" technique where models critique each other's code.
Zevi Arnovitz, a Product Manager at Meta with no formal technical background, has developed a sophisticated workflow that allows him to build and ship real applications. His process relies on treating AI as a technical co-founder and collaborator, using a structured series of steps to move from idea to production-ready code.
The Core AI-Powered Development Workflow
Zevi’s entire development process is managed within Cursor, using a series of reusable prompts called /commands that are stored within the codebase. This creates a consistent and repeatable workflow for any new feature or bug fix.
-
Issue Creation (
/create_issue): The process begins with an idea. To avoid context switching while working on another task, Zevi uses a command to quickly capture the new feature idea or bug. The AI asks a few clarifying questions and then automatically creates a detailed ticket in Linear, capturing the initial scope and requirements. -
Exploration Phase (
/exploration_phase): When ready to work on the ticket, this command prompts the AI to act as a technical lead. It pulls the context from the Linear ticket, analyzes the entire codebase to understand the current state, and identifies key areas that will be affected. The AI then asks deep, clarifying questions about the data model, UX/UI, validation, and potential edge cases. This phase is crucial for planning and avoids the common pitfall of AI agents rushing to write code without a solid plan. -
Planning (
/create_plan): After the exploration phase, the AI generates a detailed, step-by-step technical plan in a markdown file. This plan includes a TLDR, critical decisions made during the exploration, and a checklist of tasks. Having a structured plan allows for using different models for different tasks (e.g., Gemini for UI, Cursor’s Composer for speed) and serves as documentation for future work. -
Execution (
/execute_plan): With a clear plan, the AI is instructed to execute it. Zevi often uses fast models like Cursor Composer for this stage, which can implement the entire plan in minutes. -
Documentation (
/update_docs): After implementation and review, this command ensures all changes are documented, making the codebase more “AI native” and easier for agents to work with in the future.
The “Peer Review” Technique for Code Validation
As a non-technical builder, one of the biggest challenges is validating the quality and correctness of AI-generated code. Zevi developed a novel “peer review” process to address this.
-
Personifying Models: He treats different models as engineers with distinct personalities and strengths.
- Claude: The communicative, collaborative, and opinionated dev lead. Excellent for planning and learning.
- Codex (GPT): The brilliant but uncommunicative senior engineer who can solve the gnarliest bugs but offers little explanation.
- Gemini: The “crazy scientist” who is brilliant at UI and design, but whose process can be chaotic and terrifying to watch.
-
Multi-Model Review: After a feature is built (usually by Claude), he asks it to review its own code with
/review. Then, he takes the same code to Codex and Composer and asks them to perform a review as well. Each model, with its unique training and architecture, catches different types of errors. -
Consolidating Feedback (
/peer_review): Zevi then uses a final command that presents all the feedback from the other models to the original author (Claude). The prompt frames it as, “Your peers have reviewed your code and found these issues. You have the most context, so either explain why they are wrong or fix the issues.” This forces a final, critical analysis and often results in robust, well-vetted code without Zevi needing to read a single line himself.
Continuous Learning and Process Improvement
Zevi emphasizes that AI is not just for execution but also for learning.
- Learning Opportunity (
/learning_opportunity): Whenever he encounters a technical concept he doesn’t understand, he uses this command. It primes the AI to explain the topic to a “technical PM in the making” using the 80/20 rule, turning every development cycle into a learning experience. - Postmortems for Prompts: When a model makes a significant mistake, he doesn’t just fix it and move on. He asks the AI, “What in your system prompt or tooling made you make this mistake?” Based on the answer, he updates the system prompts, tooling, or documentation to ensure the mistake doesn’t happen again, creating a system that gets progressively smarter and more reliable.