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

What It Covers

Zevi Arnowitz, APM at Meta with zero coding background, demonstrates his complete workflow for building production apps using Cursor and Claude Code, including custom slash commands for planning, execution, and multi-model code review.

Key Questions Answered

• •Slash Command Workflow: Create reusable prompts in Cursor for each development phase: create-issue captures ideas to Linear mid-development, exploration-phase analyzes requirements, create-plan generates markdown templates, execute-plan builds features, review catches bugs, and peer-review uses multiple AI models to validate code quality systematically.
• •Multi-Model Code Review: Run code reviews using Claude, GPT Codex, and Gemini simultaneously, then use peer-review slash command to have Claude evaluate feedback from other models. This approach catches different bug types since each model has distinct strengths—Claude excels at communication, Codex solves complex bugs, Gemini handles UI design.
• •Progressive Exposure Therapy: Start with ChatGPT projects as a CTO copilot, graduate to Bolt or Lovable for contained features, then move to Cursor in light mode before attempting full development. This gradual progression prevents code intimidation and builds technical understanding through controlled exposure to increasing complexity levels.
• •Documentation-Driven Improvement: After bugs or failures, ask AI what in its system prompt or tooling caused the mistake, then update documentation and slash commands to prevent recurrence. This continuous postmortem process compounds learning and creates increasingly intelligent prompts that reduce future errors across the entire development workflow systematically.
• •Learning Opportunity Command: Use slash learning-opportunity prompt configured for mid-level engineering knowledge with eighty-twenty rule explanations. This transforms every confusing technical decision into a teaching moment, building genuine understanding rather than blind reliance on AI outputs, essential for maintaining code quality and making informed architectural decisions independently.