Continuous AI Prompt Improvement

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 functional products without coding knowledge. His approach combines structured prompts, multiple AI models, and continuous learning to create a complete product development system.

The workflow consists of six key phases, each with dedicated slash commands:

1. Issue Creation Phase

   • Use /create_issue to quickly capture bugs or feature ideas in Linear
   • Allows you to document ideas without breaking flow while working on other tasks
   • Creates structured tickets with TLDR, current state, and expected outcomes

2. Exploration Phase

   • Use /exploration_phase to deeply understand the problem and codebase
   • AI analyzes existing code and asks clarifying questions about requirements
   • Functions like a technical co-founder or engineering manager exploring feasibility

3. Planning Phase

   • Use /create_plan to generate a structured markdown plan document
   • Includes TLDR, critical decisions, and step-by-step tasks with status trackers
   • Creates a reference document that can be shared across different AI models

4. Execution Phase

   • Use /execute_plan to implement the planned changes
   • Different models can be assigned different parts of implementation
   • Leverages model strengths (e.g., Gemini for UI, Claude for backend logic)

5. Review Phase

   • Use /review to have the AI review its own code
   • Use /peer_review to have different AI models review each other's work
   • Creates a "team of experts" dynamic where models catch each other's mistakes

6. Documentation Phase

   • Update documentation based on what was learned
   • Ensures future code generation benefits from past experiences
   • Creates an ever-improving knowledge base for the project

Key principles for effective AI-powered development:

Continuous learning through AI

• Use /learning_opportunity when encountering unfamiliar concepts
• Have AI explain technical concepts using the 80/20 rule
• Treat each development cycle as a learning opportunity

Multi-model approach

• Use different AI models for their unique strengths:
  • Claude: Best for communication, planning, and collaborative work
  • GPT/Codex: Excellent for solving difficult technical problems
  • Gemini: Superior for UI/UX design tasks
• Have models review each other's work to catch blind spots

Systematic improvement through post-mortems

• After each bug or failure, ask AI "what in your system prompt or tooling made you make this mistake?"
• Update prompts, documentation, and workflows based on these insights
• Create an ever-improving development system that learns from past errors

Gradual progression for non-technical builders

• Start with user-friendly tools like ChatGPT projects
• Progress to semi-automated platforms like Bolt/Lovable
• Eventually graduate to direct code editing in Cursor with Claude/GPT
• Treat code exposure as "exposure therapy" - gradually increase complexity

This workflow demonstrates how non-technical product people can leverage AI to build significant products independently, while simultaneously developing their technical understanding through guided learning opportunities.