Base44 vs Cursor Tradeoffs

Zevi Arnovitz's AI-Powered Product Development Workflow for Non-Technical PMs

Zevi Arnovitz, a non-technical PM at Meta, has developed a sophisticated workflow that allows him to build complete products without coding knowledge. His approach demonstrates how AI tools can empower product managers to become builders regardless of technical background.

The workflow consists of a structured series of phases, each supported by custom slash commands in Cursor:

The Complete AI-Powered PM Workflow

1. Idea Capture and Issue Creation

• Use a "create issue" slash command to quickly document feature ideas or bugs
• The AI formats the idea into a proper Linear ticket with all necessary context
• This allows you to continue working without losing track of new ideas
• The ticket serves as a starting point when you're ready to build the feature

2. Exploration Phase

• Begin with a deep exploration of the problem before any coding
• Use the "exploration phase" slash command to analyze the issue
• The AI reads the codebase to understand technical context
• The AI asks clarifying questions about requirements, scope, and implementation
• This phase prevents premature coding and ensures proper planning

3. Planning Phase

• Use the "create plan" slash command to develop a structured implementation plan
• The plan includes a TLDR, critical decisions, and step-by-step tasks
• Each task has a status tracker that gets updated during implementation
• The plan is saved as a markdown file that becomes part of the codebase
• This documentation helps future AI interactions understand project context

4. Execution Phase

• Use the "execute plan" slash command to begin implementation
• The AI writes code according to the established plan
• Different models can be used for different aspects (e.g., Gemini for UI, Claude for backend)
• The structured plan allows for splitting work between different AI models

5. Review Phase

• Use the "review" slash command to have the AI review its own code
• Implement a "peer review" system where different AI models review each other's work
• Have Claude, GPT, and Composer each review the code independently
• Use the differences in their feedback to identify potential issues
• This creates a "team" of AI reviewers with different strengths

6. Documentation Update

• Update documentation based on what was learned during implementation
• This improves future AI interactions with the codebase
• Creates a virtuous cycle where the AI gets better at working with your specific project

Key Principles for Effective AI Collaboration

Continuous Learning and Improvement

• Use the "learning opportunity" slash command when encountering unfamiliar concepts
• Ask the AI to explain technical concepts at your level of understanding
• When AI makes mistakes, ask it to reflect on why the error occurred
• Update prompts and documentation based on these reflections
• This creates a continuous improvement cycle for both you and the AI

Model Specialization

• Different AI models have different strengths and personalities
• Claude excels at communication and planning - "the perfect CTO"
• GPT (Codex) is excellent at solving difficult technical problems - "the hoodie-wearing engineer"
• Gemini is great at UI/UX design - "the creative but sometimes chaotic designer"
• Use each model for what it does best rather than relying on just one

Progressive Technical Exposure

• Start with user-friendly interfaces like ChatGPT projects
• Graduate to more specialized tools like Bolt or Lovable
• Eventually move to direct code editing in Cursor with Claude Code
• This "exposure therapy" approach helps non-technical people become comfortable with code

AI-Native Codebase Design

• Include markdown files that explain code structure to AI
• Document key decisions and architectural patterns
• Make your codebase easy for AI to understand and navigate
• This dramatically improves the quality of AI-generated code

Broader Implications

The Changing Nature of Product Management

• "Titles are gonna collapse and responsibilities are gonna collapse"
• Everyone will become a builder regardless of technical background
• The distinction between technical and non-technical roles will blur
• The ability to use AI effectively becomes more important than traditional coding skills

Advice for Non-Technical Builders

• Start slow with user-friendly tools before diving into code
• Focus on being a "10x learner" rather than a "10x PM"
• Use AI as a learning tool, not just a production tool
• Take ownership of AI outputs - never blame the AI for poor quality
• Treat AI interactions as "tuition" - an investment in your learning