Lenny's PodcastMeta PM Zevi Arnovitz: How a non-coder ships real features
Through staged Cursor workflows and multi-model peer review on Claude; Codex and Gemini help a non-technical PM turn Linear tickets into shipped features.
CHAPTERS
- 0:00 – 1:11
Hook: a non-technical PM shipping real product with AI coding tools
Lenny tees up the central premise: Zevi ships production software despite not writing or reviewing code traditionally. Zevi frames AI as “superpowers” that will turn more people into builders and collapse traditional role boundaries.
- •Zevi has “zero technical background” yet ships product
- •Cursor + Claude Code as the current tool stack
- •Code review is the hardest part for non-technical builders
- •AI advantage: you’re replaced by people who use AI better
- •Encouragement for juniors to build independently
- 1:11 – 6:05
Who Zevi is and what listeners will get (prompts, commands, workflow)
Lenny introduces Zevi (Meta PM, previously Wix) and sets expectations for a highly tactical episode. He highlights downloadable prompts and commands that replicate Zevi’s end-to-end workflow—from ideation to documentation.
- •Meta PM with hands-on “vibe coding” workflow
- •Show-notes download: prompts + slash commands
- •Workflow covers Linear → planning → build → review → docs
- •Goal: make non-technical product people effective builders
- 6:05 – 7:42
Zevi’s origin story: from music student to “AI superpowers” in Japan
Zevi shares how discovering tools like Bolt/Lovable around Sonnet 3.5 triggered his shift into building. He emphasizes the episode’s success metric: listeners starting to build, not admiring him.
- •No army tech unit; studied music/psychology
- •Saw AI app-building demos (Bolt/Lovable) and started immediately
- •Motivation: build consistently over the past year
- •Claude’s framing: success = audience starts building
- 7:42 – 13:19
Why projects matter: compartmentalizing context and creating an AI “CTO”
Zevi explains how ChatGPT/Claude Projects solved context-mixing issues from AI memory and enabled a dedicated “CTO” persona. He describes prompting the CTO to challenge him and avoid sycophantic behavior.
- •Projects = shared instructions + shared knowledge base
- •Memory can mix unrelated contexts (running vs. PM work)
- •CTO prompt: AI owns technical decisions, challenges assumptions
- •Sycophancy pitfalls (e.g., AI “riffing” wrong technical claims)
- •Recommendation: start in chat before moving into code tools
- 13:19 – 14:47
Graduating tools: Bolt/Lovable → Cursor, and why control beats “opinionated” scaffolding
Zevi describes outgrowing early vibe-coding tools when complexity increases (e.g., payments, data changes). Cursor/Claude Code gives more direct access and control, but requires more decision-making and learning.
- •Early tools are eager to code; planning becomes critical later
- •Complex changes (payments/db) punish “just start coding” behavior
- •Cursor = models + fewer guardrails; more control for advanced work
- •Advice: gradual exposure (projects → Bolt/Lovable → Cursor)
- 14:47 – 17:22
Screenshare setup: Cursor, Claude Code, and reusable slash commands
Zevi reveals his practical system: saved slash commands as reusable prompts stored with the codebase. He walks through the command set that drives his workflow from issue capture through docs updates.
- •Slash commands = reusable prompts invoked via “/command”
- •Core flow: Create Issue → Exploration → Create Plan → Execute → Review → Peer Review → Update Docs
- •Commands are designed for speed while “mid-development”
- •Claude can access the full repo context inside Cursor
- 17:22 – 20:44
Product demo: StudyMate and the feature to build (fill-in-the-blank questions)
Zevi demos StudyMate, a weekend side project that generates quizzes from uploaded study materials. He chooses a competitor-inspired feature to implement live: fill-in-the-blank questions with drag-and-drop answers.
- •StudyMate: upload materials → generate interactive quizzes
- •Current format: multiple choice only
- •New feature: 30% fill-in-the-blank questions; drag-and-drop UI
- •Answer design: two blanks with multiple candidate choices
- 20:44 – 23:14
From idea to ticket: capturing a Linear issue via MCP tools
Zevi dictates requirements and uses his “Create Issue” command to generate a structured Linear ticket. He explains MCP (Anthropic’s tool-use framework) powering direct actions like writing issues to Linear.
- •Voice dictation to describe requirements quickly
- •Create Issue command asks minimal clarifying questions
- •MCP enables Claude to create the Linear ticket automatically
- •Ticket is a starting point for exploration—not a final spec
- 23:14 – 29:18
Exploration phase: repo-aware clarification before coding
Zevi runs “Exploration Phase” against the Linear ticket, prompting Claude to read code, map affected areas, and ask high-leverage questions. He also introduces “Learning Opportunity” to build technical intuition using 80/20 explanations.
- •Exploration pulls context from Linear (e.g., STU-88)
- •Claude reads codebase, summarizes current architecture and constraints
- •Clarifying questions cover scope, UX, validation, grading, prompt changes
- •Learning Opportunity command: explain concepts at the right level
- 29:18 – 30:49
Planning into a markdown build plan (and why plans enable multi-model work)
Using “Create Plan,” Zevi generates a structured markdown plan with tasks and status tracking. He explains how a written plan lets him delegate parts to different models (e.g., Gemini for UI) while keeping execution coherent.
- •Plan template: TL;DR, critical decisions, step-by-step tasks with status
- •Markdown plan lives in repo for future agents and continuity
- •Model specialization: fast Composer for routine work; Gemini for UI
- •Plans reduce chaos and keep execution aligned
- 30:49 – 38:33
Executing in minutes: Cursor Composer implements the feature, then local QA
Zevi triggers execution of the plan in Cursor, emphasizing speed and flow. He runs the app locally to validate behavior before moving into deeper review.
- •Composer executes plan rapidly to produce working code changes
- •Work happens on a branch, not directly on production
- •Local run + manual QA catches obvious behavioral issues
- •AI spend treated as “tuition” for speed + learning
- 38:33 – 40:39
Code review without being technical: self-review + multi-model reviews + peer review arbitration
Zevi addresses the core bottleneck: trusting and reviewing AI-generated code. He runs Claude’s self-review, then gets independent reviews from other models (Codex, Composer), and uses a “Peer Review” prompt to arbitrate conflicts and drive fixes.
- •Manual QA first, then automated review (/review)
- •Multiple model reviews catch different classes of issues
- •Peer Review prompt: treat Claude as dev lead responding to other leads’ feedback
- •Repeat until issues converge; use Learning Opportunity to understand fixes
- 40:39 – 45:40
Personifying models: choosing the right AI for the job
Zevi explains his mental model for model behavior differences and how that informs tool choice. He characterizes Claude as a collaborative CTO, Codex as a terse but elite bug-fixer, and Gemini as a risky-but-brilliant designer.
- •Claude: communicative, opinionated, collaborative “CTO/dev lead”
- •Codex: strong coding/debugging, low communication
- •Gemini: excellent design output, messy execution path
- •Strategy: play to strengths, mitigate weaknesses via cross-review
- 45:40 – 51:05
Postmortems and “making the codebase AI-native” through docs and tooling
Zevi argues that the biggest productivity gains come from systematic iteration: when AI fails, he diagnoses root causes and updates prompts, documentation, and tooling so mistakes don’t repeat. This turns the workflow into a compounding asset.
- •Do postmortems on AI mistakes (not just “try again until it works”)
- •Ask: what in prompts/tooling caused the failure?
- •Update docs, instructions, and reusable commands to prevent repeats
- •Treat documentation as leverage for future agent performance
- 51:05 – 53:43
Using this at larger companies: safe scopes, PRs, and AI-readability of repos
Zevi discusses how parts of this workflow can translate to bigger orgs, with guardrails. The key is making repositories navigable for agents (clear markdown, conventions) and limiting PM-coded changes to contained, reviewable work.
- •Step 1: make the codebase “AI-native” with human-led structure/docs
- •PMs shouldn’t ship risky migrations; start with contained UI changes
- •Submit PRs for engineers to finalize and validate
- •Selling internally: skeptics exist; teams that invest in setup get ahead
- 53:43 – 58:12
PM craft and output quality: avoiding “AI slop” with context and ownership
Zevi rejects the idea that AI inherently weakens PM thinking—misuse does. He emphasizes owning outputs, providing context and style constraints, and using AI as a mentor/accelerator rather than a content generator.
- •AI isn’t “outsourcing thinking” if you still own decisions
- •Bad outcome = publishing unreviewed AI output and blaming the tool
- •Improve quality by supplying constraints, context, and your writing style
- •Mentions emerging “deslop” concepts for cleaning outputs
- 58:12 – 1:02:48
AI-assisted Meta interviewing: projects as a coach, question mining, and feedback loops
Zevi outlines how he used a dedicated Claude project as an interview coach, collected best frameworks, and ran mock interviews. He also used question-bank analysis to prioritize practice, while emphasizing that human mocks still matter most.
- •Create an interview-coach project with curated frameworks/resources
- •Mock interviews with AI + strong feedback prompts (no sycophancy)
- •Analyze real question banks to prioritize what to practice
- •Use AI to generate “ideal answers” for learning patterns
- •Human mocks remain essential for top-tier performance
- 1:02:48 – 1:07:44
Failure Corner: early Wix mistake—trying to impress vs. becoming a 10x learner
Zevi shares a formative failure: an underwhelming product review after working solo to ‘wow’ the team. He reframed success as learning, sought targeted mentorship, and improved by leveraging others’ strengths.
- •Initial approach: hide work, grind alone, aim to impress
- •Result: missed expectations, unanswered questions, weak review
- •Reframe: no one expects a 10x PM; they expect a 10x learner
- •Deliberately mentor-map colleagues by strengths (sense, frameworks, systems thinking)
- 1:07:44 – 1:13:29
Lightning round: recommendations, mottos, and early entrepreneurial stories
Zevi shares favorite books, shows, products, and personal mottos that match his builder mindset. He closes with a high-school entrepreneurship story showcasing creative distribution and marketing instincts.
- •Books: The Fountainhead, Shoe Dog, Mindset
- •Shows: Severance and The Pitt
- •Product: Cap (open-source Loom alternative)
- •Mottos: “You can just do things” and “Nobody knows what they’re doing”
- •Thermal clothing hustle: distribution + chant marketing
- 1:13:29 – 1:15:12
Where to find Zevi, how to engage, and closing message: go build
Zevi invites people to reach out and to try his products for feedback. Lenny reiterates the value of the workflow and the goal of helping people cross the intimidation barrier into building.
- •Reach Zevi on LinkedIn/X; he’s open to helping others
- •Try StudyMate (students) and Dibur2text (Israel) and share feedback
- •Core takeaway: start small, learn fast, and build consistently
- •Episode closes with calls to subscribe/review
