All podcasts / Lenny's Podcast / Summary

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

2026-01-18 - 75 min - source - Read full transcript
Lenny Rachitsky (host)Zevi Arnovitz

Key insights

A non-technical PM can ship real, revenue-generating products by running every project through a fixed six-phase slash-command pipeline: capture an issue, explore it with AI, write a plan, execute, review, and update documentation.
Zevi built this pipeline as reusable saved prompts inside Cursor and Claude Code over about a year, starting from zero coding background. He is releasing the actual prompt files so listeners can copy the workflow directly rather than reconstruct it.
ai-native-building
Giving an LLM an explicit 'CTO' persona, told to own decisions, push back, and never be a people-pleaser, counteracts the default sycophancy that otherwise validates technically wrong ideas.
Zevi recounts ChatGPT confidently claiming two unrelated technologies (Bun and Zustand) were 'exactly the same' and then admitting it was 'just making things up.' A dedicated project with a strict, opinionated system prompt largely fixes this failure mode.
ai-native-building
Different frontier models have distinct practical strengths worth routing tasks to deliberately, rather than using one model for everything.
He splits work by model: Claude for planning and communication, Cursor's Composer for speed on simpler tasks, Gemini 3 for UI and visual design despite an alarming-looking process, and Codex for isolated, hard bugs it works on alone for hours before returning a fix.
multi-model-workflow
The hardest problem once AI makes code cheap to write is verifying it, which Zevi solves with a 'peer review' step where multiple models review the same code and are told to contest each other's findings rather than have changes accepted automatically.
He runs Claude, Codex, and Composer as independent reviewers on the same branch, then feeds each model's findings back to the original coding agent as 'dev lead one' and 'dev lead two,' instructing it to either justify why a flagged issue is not real or fix it.
ai-code-review
Beginners should ease into AI coding tools gradually, from a conversational chat project through no-code builders to a full IDE, because a slow on-ramp builds real understanding instead of blind trust in code they cannot read.
His recommended progression is a ChatGPT/Claude project, then Bolt or Lovable, then Cursor in light mode, then eventually a full terminal-based dev setup. He frames this as exposure therapy for people who find code 'terrifying.'
ai-native-building
Opinionated no-code builders like Bolt, Lovable, Replit, Base44, and v0 remove hard technical decisions to make building easy, but that same abstraction removes control once a builder wants to make deliberate architecture choices.
Zevi says all these tools run essentially the same underlying models as Cursor and Claude Code, but add layers that pre-decide things like database and auth setup. He outgrew Bolt specifically when he needed to add payments and wanted more control than the tool allowed.
ai-native-building
Post-mortems that turn every AI mistake into a permanent fix to the prompt or documentation compound over time into a workflow that keeps getting smarter, rather than repeating the same failure.
When Claude fails or produces a bad result, Zevi asks it to explain what in its own system prompt or tooling caused the mistake, then updates the relevant slash command or docs so future runs avoid it, treating each error as a root-cause fix rather than a one-off retry.
ai-native-building
Using AI extensively is not 'outsourcing your thinking' - the PM job was never about having the smartest answer alone, it is about getting the team to the right solution fastest.
Zevi pushes back hard on colleagues who called his AI co-pilot use outsourcing, arguing that discomfort with AI assistance often correlates with people who also don't like sharing unfinished work; he says owning the output, not who or what produced the first draft, is what matters.
pm-career-mindset
AI-assisted interview prep, including a personal Claude 'coach' project, a custom quiz game built in Base44 to drill product-segmentation questions, and scraping a real interview question bank via Perplexity's Comet browser, still cannot fully replace human mock interviews for something as competitive as a Meta PM loop.
Zevi used AI heavily to prepare for his Meta interviews but says the single biggest unlock was cold-outreaching real people on LinkedIn for live mock interviews, because AI mocks 'take you to a point' but human feedback catches things AI still misses.
pm-career-mindset
Junior employees should optimize for being the best learner in the room rather than trying to prove expertise immediately, since managers expect learning speed, not immediate mastery.
After bombing his first product review at Wix by working alone to try to impress a highly experienced PM team, Zevi explicitly recruited each senior PM as a mentor for their specific strength (product sense, methodology, systems thinking). This turned his later successes into shared wins with his mentors rather than solo flexes.
pm-career-mindset
As AI narrows the gap between technical and non-technical roles, Zevi expects job titles and rigid responsibilities to 'collapse,' making this an unusually good time to be a junior or self-taught builder despite a difficult formal job market.
He argues that curious, hardworking, kind communicators now have outsized leverage compared to people with many years of experience but less fluency with AI tools, since the barrier to building and shipping has dropped so far.
pm-career-mindset
Making a codebase 'AI-native' - plain-language markdown files explaining architecture and conventions for agents to read - is a prerequisite technical people should build before non-technical staff attempt to ship contained changes at a larger company.
Zevi says he still would not expect PMs to touch database migrations or large systems, but with the right AI-native documentation in place, contained UI changes handed off to an engineer for final polish are realistic today and will become more common.
ai-native-building

Books referenced

Media referenced

Companies

Techniques and frameworks

Summary

Zevi Arnovitz, a product manager at Meta with no technical or engineering background, walks Lenny through the exact workflow he built over the past year to design, build, and ship real software using Cursor and Claude Code, including a side app called StudyMate that makes money. The conversation opens with his origin story: watching a YouTube demo of Bolt or Lovable while traveling in Japan around the time Claude 3.5 Sonnet launched, feeling like he'd been handed "superpowers," and going home to start building immediately. From there he built up, iteration by iteration, a six-phase pipeline of reusable slash-command prompts inside Cursor: capture an idea as a Linear issue, explore it with AI, write a plan, execute the plan, review the resulting code, run a multi-model peer review, and update documentation, all of which he is releasing publicly alongside the episode.

A recurring thread is how Zevi manages the model itself. He originally built a dedicated ChatGPT "CTO" project with a strict system prompt demanding ownership, pushback, and zero sycophancy, after watching ChatGPT confidently fabricate a false technical claim and then admit it was "just making things up." That persona now lives inside his Claude Code setup. He also personifies different models by their practical character: Claude as an opinionated but collaborative CEO-type collaborator, Codex as a heads-down solo engineer who disappears for hours and returns with a fix, and Gemini 3 as a brilliant but visibly erratic UI designer. He deliberately routes tasks to whichever model fits, using Cursor's fast Composer model for simple work and splitting front-end from back-end work across models.

The most detailed technical segment is a live screen-share where Zevi builds a real feature for StudyMate (fill-in-the-blank quiz questions) end to end, narrating each slash command as he goes. He argues that writing code is no longer the hard part; reviewing AI-written code is, since he personally cannot catch subtle bugs by reading code he doesn't understand. His answer is a "peer review" step where Claude, Codex, and Composer each independently review the same branch, and their findings are fed back to the original coding agent as competing "dev lead" opinions it must either justify overriding or act on. He treats every mistake as a signal to update the underlying prompt or documentation so the workflow compounds in quality over time, rather than repeating the same failure.

Beyond the technical mechanics, Zevi and Lenny spend real time on what this means for the PM role and for careers more broadly. Zevi rejects the framing that heavy AI use is "outsourcing your thinking," arguing the PM's job was never to have the smartest answer solo, and that resistance to AI assistance often correlates with discomfort sharing unfinished work. He also describes using AI extensively, but not exclusively, to prepare for his Meta PM interviews: a personal Claude "coach" project, a custom Base44 quiz game to drill product-segmentation questions, and scraping a real interview question bank via Perplexity's Comet browser, while stressing that live human mock interviews were still the single biggest unlock for a loop as competitive as Meta's.

The episode's Failure Corner covers Zevi's rocky start at Wix, where he tried to impress a team of far more experienced PMs solo, bombed his first product review, and learned to instead recruit each senior PM as a mentor for a specific strength, turning later wins into shared successes rather than solo flexes. He closes on a broader claim: as AI narrows the gap between technical and non-technical roles, job titles and rigid responsibilities will "collapse," and curious, hardworking people willing to learn fast have outsized leverage right now, regardless of formal experience. A lightning round covers his book picks (The Fountainhead, Shoe Dog, Mindset), TV recommendations (The Pitt, Severance), and a story about his teenage thermal-clothing side hustle, complete with a basketball-arena marketing jingle.

Notable Quotes

"If people walk away thinking how amazing you are, you failed. And if people walk away and open their computer and start building, you've succeeded." - Zevi Arnovitz, quoting advice from Claude

"It's not that you will be replaced by AI. You'll be replaced by someone who's better at using AI than you." - Zevi Arnovitz

"Code is just words at the end of the day. So it's just files on your computer." - Zevi Arnovitz

"The only way that AI makes you worse at your job is if you're using it wrong." - Zevi Arnovitz

"Nobody knows what the fuck they're doing." - Zevi Arnovitz, citing his brother's life motto