A non-technical Product Manager building products with AI

In a recent Lenny's Podcast, Zevi Arnovitz, a product manager at Meta, describes how he ships real features without writing code. He has no technical background - he studied music in high school and wasn't in a tech unit during his army service. A year ago, while travelling in Japan, he watched a YouTube video about building apps with AI and thought: "Someone just told me I have superpowers."

Since then, he's built StudyMate, a side project that generates interactive quizzes from uploaded study materials. He's made it profitable. He's localised it from Hebrew to English in two days. He's built a personal website from scratch in 90 minutes. All without writing a line of code himself.

His workflow centres on Cursor with Claude Code, but he started much simpler. He began with a ChatGPT project that acted as his CTO - a technical co-founder who would challenge his thinking rather than just agree with everything. This taught him to think about architecture and technical decisions before graduating to coding tools.

The workflow: six commands

Zevi's process breaks down into six reusable prompts he calls /commands. These live in his codebase and inject specific instructions into Claude at each stage:

/create-issue captures ideas quickly without breaking flow. When he's mid-development and spots a bug or has a feature idea, Claude asks brief questions and creates a ticket in Linear using MCP (Model Context Protocol), Anthropic's tool that gives AI the ability to interact with external services.

/exploration is where planning happens. Claude reads the Linear ticket, analyses the codebase, and asks clarifying questions about scope, data models, UI decisions, validation logic, and system prompts. This stage takes time - Zevi doesn't rush it. The AI needs to deeply understand both the problem and the current state of the code.

/create-plan outputs a markdown file with clear, minimal steps. Each task has a status tracker that updates as work progresses. The plan includes a TL:DR and critical decisions. Having this as a separate file means Zevi can split work across different models - Cursor's Composer for backend speed, Gemini 3 for frontend design.

/execute is where the code gets written. Cursor's Composer is fast enough that entire features take minutes rather than days.

/review tells Claude to review its own code. But Zevi doesn't stop there. He opens multiple AI models—Claude Code, Codex (ChatGPT's coding tool), and Cursor - and has each review the same code independently. Then comes the clever bit.

/peer-review frames Claude as the dev lead whose code other team leads have reviewed. The prompt explicitly tells Claude not to take their feedback at face value - it has more context and led the project, so it should either explain why the issues aren't real or fix them. Sometimes Claude gets defensive: "This has been raised for the third time. For the third time, this is not an issue. This is by design."

Three models, three personalities

Zevi thinks about AI models as people. Claude is the perfect CTO - communicative, smart, opinionated but collaborative. Codex is the best coder in the company who wears a hoodie and sandals, sits in a dark room, and only emerges to say "I fixed it" without explanation. Gemini is a brilliant but chaotic designer whose thought process is terrifying to watch but produces beautiful results.

He plays to their strengths. Claude handles planning and execution. Codex tackles gnarly bugs. Gemini does UI work. And when they disagree during peer review, he lets them fight it out until the issues are resolved.

Learning, not just building

Zevi has another command: /learning-opportunity. When something's difficult to understand, this prompt tells Claude he's a technical PM in the making with mid-level engineering knowledge. Claude then explains using the 80/20 rule - enough detail to understand without drowning in complexity.

After shipping features, he runs post mortems. When Claude makes mistakes, he asks what in its system prompt or tooling caused the error. Then he updates documentation so the mistake never happens again. The models get smarter, and so does the workflow.

Starting small

His advice for non-technical people: don't jump straight to Cursor. Start with a ChatGPT project. Tell it your idea. Ask it to explain the first steps. Be inquisitive. Learn. Think of it as exposure therapy for code - gradually ease in until you're comfortable opening a terminal in dark mode.

Products like Bolt, Lovable, and Replit make excellent stepping stones. They're opinionated, which means less control but also less decision-making. They handle authentication and databases automatically. Once you outgrow them, move to Cursor.

The shift for PMs

Zevi used AI to prepare for his Meta interview. He created a project that acted as his coach, fed it frameworks and advice from experienced PMs, ran mock interviews, and had Claude give him brutal feedback. He built a quiz game in Base44 to practice product segmentation questions on the bus.

He sees AI as a learning accelerator, not a shortcut. When he started at Wix, he tried to impress everyone in his first product review by working alone for hours. He failed. His manager's expectation wasn't that he'd be a 10X PM - it was that he'd be a 10X learner. That shift in mindset changed everything.

The fear that AI makes you worse at your job only holds if you use it wrong. If you treat it like a thought partner - something that helps you think through problems faster and learn from the process, it makes you significantly better.

Code is just words. Files on a computer. And now those words can be written by something that's always available, doesn't judge, and can explain its reasoning. For curious, hardworking people, this is the best time to be building.

If you want the full AI development workflow: slash commands, CTO prompt, and step-by-step examples from the podcast, visit Zevi's website.