LLenny's Podcast
← All episodes
Zevi Arnovitz (Meta)18 January 2026

The non-technical PM’s guide to building with Cursor

8Frameworks
14Insights

Frameworks in this episode

Innovation4 steps

Adversarial Multi-Model Peer Review

Have rival LLMs review each other's code and fight it out until no issues remain

Read →
Strategy4 steps

Model Casting by Personality

Route each task to the model whose 'personality' matches, and split plans to fit

Read →
Productivity4 steps

Root-Cause Tooling Post-Mortem

When AI botches something, ask what in its prompt caused it — then patch the tooling

Read →
Innovation3 steps

The AI-as-CTO Persona Project

Cast the AI as an opinionated technical co-founder to kill sycophancy and premature coding

Read →
Peak Performance5 steps

The AI Interview-Prep Coach System

Build an AI coach, mine the real question bank, drill weak spots — then mock with humans

Read →
Self-Mastery4 steps

The Exposure-Therapy Tool On-Ramp

Ease into coding tools in graduated steps so code stops being terrifying

Read →
Self-Mastery3 steps

The Learning-Opportunity Command

Turn every confusing moment into an 80/20 lesson aimed at your current skill level

Read →
Productivity7 steps

The Seven-Command Build Loop

Six slash-commands turn vibe-coding into disciplined software delivery for non-coders

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 4

Hot Take10:00

Why Plain ChatGPT Would Be the Worst CTO

Zevy argues that default ChatGPT is too much of a people-pleaser to trust with technical decisions. He tells a story about asking whether Bun JavaScript was similar to Zustand (two unrelated things), and ChatGPT agreed enthusiastically, then admitted it thought he was just making things up and riffing. This sycophancy is why he built a dedicated 'CTO' project with a prompt that challenges him instead of agreeing.

  • Default ChatGPT agrees with your dumbest ideas rather than pushing back
  • It falsely claimed Bun JavaScript and Zustand were 'exactly the same'
  • When corrected, it said it thought he was making it up and riffing
  • The fix: a project-scoped CTO prompt told to challenge, not people-please

“he goes oh I'm Sorry, I thought you were just making this up and I was riffing with you”

Zevy Arnovitz · 11:00

“if regular Chachi PT was a CTO, that would be the CTO who like goes along with your dumbest ideas.”

Zevy Arnovitz · 11:30
#sycophancy#chatgpt#ai-behavior#prompting
Hot Take40:30

Claude, Codex, and Gemini as Three Very Different Coworkers

Zevy imagines each coding model as a distinct person with distinct characteristics. Claude is the perfect CTO: communicative, smart, opinionated, and collaborative. Codex is the elite coder in a hoodie and sandals in a dark room you only bother for the worst bugs. Gemini is a brilliant but terrifying artsy scientist whose chaotic process ends in beautiful design. Playing to each model's strengths and covering their weaknesses with other models is his game changer.

  • Claude reads as a communicative, opinionated, collaborative CTO
  • Codex is the antisocial elite coder you only call for the worst bugs
  • Gemini's coding process is scary to watch but produces great design
  • Using multiple models to offset each other's weaknesses is a game changer

“the best coder within the company who comes to the office like with uh a hoodie and sandals and sits in a dark room and…”

Zevy Arnovitz · 41:30

“Gemini is like a crazy scientist who's super artsy, uh, super talented at designing, but if you sit next to it and watch it work,…”

Zevy Arnovitz · 42:00
#models#claude#codex#gemini
Hot Take54:00

AI Slop Is Human Error: Own Your Outputs

Zevy strongly disagrees with the idea that using AI is 'outsourcing your thinking' or that it produces slop. He argues the belief that a PM's job is always having the right answers and being the smartest person in the room is a misconception, and that AI is like an always-available, non-judgmental mentor. If you ship AI output without owning it, that's your mistake, not the AI's, and used intentionally AI lets junior PMs play at a much higher level.

  • 'Outsourcing your thinking' is the worst way to frame using AI
  • The idea that a PM must always have the answers is a misconception
  • Publishing AI output you don't own is human error, not AI slop
  • AI lets junior PMs operate at a strategic level and get more reps

“there's a misconception with a lot of PMs that the job is always having the right answers and being the smartest person in the room.”

Zevy Arnovitz · 55:00

“the only way that AI makes you worse at your job is if you're using it wrong.”

Zevy Arnovitz · 57:00
#ai-slop#pm-craft#junior-pms#ownership
Hot Take1:06:30

It's the Best Time in History to Be a Junior

Countering the narrative that junior roles are disappearing, Zevy argues it's actually the best time ever to be a junior or a learner, because never before could you leave school and bootstrap a startup on your own. He says a curious, hardworking, kind person who communicates well now has an unfair advantage and can deliver more value than many people with 20 years of experience.

  • Yes, traditional junior roles are shrinking, but the opportunity has grown
  • You can now leave school and bootstrap a startup on your own
  • He saw more and more people building their own things with AI
  • Curious, kind, communicative people have an unfair advantage

“it's the best time to be a junior. It's the best time to be a learner.”

Zevy Arnovitz · 1:07:00

“if you're a kind person and a good communicator, you have such an unfair advantage and you can give more value to companies than most…”

Zevy Arnovitz · 1:07:00
#career#juniors#ai-opportunity#optimism

Explainer· 1

Explainer32:00

The Real Difference Between AI Coding Tools Is the Harness, Not the Model

Zevy explains that Bolt, Lovable, Cursor, and Claude Code often run the same underlying models, so the differentiator is the harness around them. Bolt and Lovable add layers that take guesswork and hard decisions off the user's plate, which makes building easy but removes control. Claude Code, by contrast, drops the model straight into your codebase with full tools and hands you all the decisions.

  • The models are largely the same across tools; the harness differs
  • Bolt and Lovable add middle layers that remove hard decisions for the user
  • Easier to build, but you get less control in exchange
  • Claude Code gives the model full tools in your codebase, plus all the decisions

“the main difference between all these tools is basically the harness. So, the models are all the same models.”

Zevy Arnovitz · 32:00

“claude code is just taking claude and shoving it straight in your code system and giving it full tools and to do whatever it wants.”

Zevy Arnovitz · 32:30
#ai-tools#harness#cursor#trade-offs

Story· 5

Story06:00

The Sonnet 3.5 Moment That Made a Non-Coder Feel Like He Had Superpowers

Zevy, a self-described zero-technical-background PM, recounts watching a YouTube video in Japan when Sonnet 3.5 came out, showing apps being built with Bolt or Lovable. The moment felt like being handed superpowers, and the second he got home he ran to his computer without unpacking. He frames the whole episode's goal around a line Claude gave him while prepping.

  • He has zero technical background and did music in high school
  • The trigger was a Greg Eisenberg or Riley Brown YouTube video during a trip to Japan
  • He opened Bolt the moment he got home and has been building for a year since
  • His goal for the episode: inspire people to open their computer and build, not just admire him

“it basically felt like someone came up to me and said, "Hey, Zevie, there's this cool new technology you should check out. Um, you should…”

Zevy Arnovitz · 06:30

“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.”

Zevy Arnovitz · 07:00
#origin-story#ai-coding#non-technical#inspiration
Story35:00

Time Machine Moments: Three AI Agents Working in Parallel

Zevy describes what he calls 'time machine moments' where he has so much AI leverage he runs out of things to do. In one week he localized Studymate from Hebrew to English in two days, built a personal site from nothing to live in under two hours, and ran all three projects in parallel, with nothing to do but let the agents think.

  • Localized Studymate Hebrew-to-English in two days, work he estimates would take a dev team weeks
  • Built and shipped a personal site to a live domain in about 90 minutes
  • Ran all three projects simultaneously with agents thinking in parallel
  • He calls these 'time machine moments' of feeling like he's living in the future

“I was fully localizing uh studymate from Hebrew to English which I did in two days which would probably take a dev team weeks”

Zevy Arnovitz · 35:00

“I'll just say we we live in the future.”

Zevy Arnovitz · 35:30
#productivity#parallel-agents#leverage#vibe-coding
Story58:30

How He Used AI (and Human Mocks) to Land the Meta PM Job

When Meta reached out, Zevy immediately built a Claude 'coach' project fed with the best interview frameworks online, including Ben Erez's material, and built a Base44 quiz game to drill product segmentation on his bus commute. He had Claude mock-interview him, play the perfect candidate so he could learn from a great answer, and analyzed a free question bank via Perplexity's Comet browser to prioritize prep. But he says the biggest game changer was cold-outreaching people on LinkedIn for real human mocks.

  • Built a Claude coach project loaded with the best interview frameworks
  • Built a Base44 quiz game to practice product segmentation on the bus
  • Had Claude play the perfect candidate so he could learn from ideal answers
  • The biggest game changer was cold-outreaching on LinkedIn for human mocks

“the biggest game changer for me was doing uh human mocks. So cold outreaching to people uh on LinkedIn and having them uh do actual…”

Zevy Arnovitz · 1:00:00

“you're my coach, and I don't want you to make me feel good. I want you to make me, um, as ready as possible for…”

Zevy Arnovitz · 1:02:00
#interview-prep#meta#ai-coach#base44
Story1:03:30

The Wix Failure That Taught Him to Be a 10x Learner

In failure corner, Zevy tells how he started at Wix on the elite editor team and, trying to impress four far-more-experienced PMs, worked alone in secret and failed his first product review badly, in the wrong format and full of gaps he'd missed. The lesson: they expected a 10x learner, not a 10x PM. He then mapped each teammate's strength (product sense, methodology, systems thinking) and used them as mentors, which made their next success feel like a shared win.

  • He worked alone to impress senior PMs and failed his first product review
  • The real expectation was to be a 10x learner, not a 10x PM
  • He mapped each teammate's strength and used them as targeted mentors
  • Asking for help turned his success into the team's shared success

“they had zero expectation of me being a 10x PM, but the expectation of me was being a 10x learner.”

Zevy Arnovitz · 1:04:30

“be the best learner you can be at the beginning. No one expects you to know all the answers and no one expects you to…”

Zevy Arnovitz · 1:04:30
#failure#learning#mentorship#junior-pm
Story1:11:00

The High School Thermal Clothing Hustle (Pre-ChatGPT Edition)

Zevy tells the story of selling thermal clothing in 10th grade in Jerusalem, where he was sixth or seventh down the supply chain making only about $4 a sale. Over a summer he negotiated directly with the importer, stalling on the phone while Googling import taxes to counter him, and landed a price that gave him roughly 100% profit. He then recruited the coolest kids across schools to sell and even wrote a drum-backed basketball chant with his phone number in it, so people in Jerusalem still know his number by the tune.

  • As a middleman he made only about $4 per sale, six or seven links down the chain
  • He negotiated directly with the importer over a whole summer
  • He Googled import taxes live on calls to counter the importer's objections
  • He got roughly 100% margin, recruited sellers, and wrote a viral basketball chant with his number

“I said listen man I'm finishing school soon. This is not going to be my career. Either do it or not.”

Zevy Arnovitz · 1:11:30

“sometimes when I walk in Jerusalem people stop me and say like hey it's thermals heavy”

Zevy Arnovitz · 1:13:00
#entrepreneurship#negotiation#marketing#origin-story

Q&A· 1

Q&A51:30

How Much of This Can a PM Actually Ship at a Larger Company?

Asked how this translates to a 500-1000 person company, Zevy says the first step is making the codebase AI-native with plain-text markdown docs that explain to agents how to work in each area, and that this must be done by technical people. He still doesn't think PMs should ship heavy database migrations or big projects, but contained UI work, especially building it, opening a PR, and handing it to a dev for final touches, is realistic and coming.

  • Step one is making the codebase AI-native with markdown docs for agents
  • That AI-native setup should be done by technical people
  • PMs should not ship heavy database migrations or big projects
  • Contained UI work plus a PR handed to a dev for finishing is viable

“first making your codebase AI native is a really important step and I think this needs to be done by technical people.”

Zevy Arnovitz · 51:30

“I still don't think uh like PMS should be shipping heavy uh database chain migrations or any like big project.”

Zevy Arnovitz · 52:00
#pm-role#enterprise#ai-native#codebase

Takeaway· 3

Takeaway12:30

Treat Code Like Exposure Therapy: Ease In Gradually

For non-technical people, Zevy says code is genuinely terrifying, so he recommends easing in like exposure therapy rather than jumping straight to a scary dev environment. Start with a beautiful, simple ChatGPT project, graduate to Bolt or Lovable, then Cursor in light mode, and only eventually open a terminal and go full dark-mode dev.

  • Non-technical people find code the scariest thing to look at
  • Start with a ChatGPT project because it keeps you conversing, not coding
  • Progression: GPT project -> Bolt/Lovable -> Cursor light mode -> terminal/dark mode
  • The point is to take time to converse and learn, not rush into raw code

“if you're nontechnical like me, code is terrifying. It's the scariest thing in the world to look at. And I look at it as kind…”

Zevy Arnovitz · 12:30

“go to cursor in light mode slowly slowly gradually ease in until you like open a terminal uh you know go full dark mode go…”

Zevy Arnovitz · 13:00
#onboarding#learning#non-technical#cursor
Takeaway14:00

Code Is Just Words, So You Can Carry a Project From App to App

Zevy shares a reframe he got from Tal Riv: code is just words, which means it's just files on your computer. Because of that, you can work on the same project across different tools and even run multiple models and apps on it at once, moving fluidly as you outgrow each tool.

  • Code is just words, i.e. just files on your computer
  • You can carry the same project from app to app
  • You can work with multiple models and apps on one project
  • He graduated tools as he outgrew them (Bolt broke when connecting payments)

“code is just words at the end of the day. Um, so it's just files on your computer.”

Zevy Arnovitz · 14:00
#mental-model#tooling#workflow#cursor
Takeaway46:00

When the AI Fails, Ask It What in Its Prompt Caused the Mistake

Zevy's biggest productivity hack is constant post-mortems. Instead of just hammering at a problem until it works, when Claude fails or reveals a misunderstanding he asks what in its system prompt or tooling caused the mistake, has it introspect, and then updates the documentation or tooling so that mistake never recurs. He treats iterating on prompts and watching responses improve as the dividing line between people who are okay with AI and people who truly know how to use it.

  • Do constant post-mortems instead of running at the wall until it works
  • Ask the model what in its system prompt or tooling caused the mistake
  • Update docs and tooling so the same mistake can't happen again
  • Iterating on prompts is what separates real AI users from casual ones

“updating documentation and tooling is one of the biggest hacks for productivity.”

Zevy Arnovitz · 46:00

“I'll ask it what in your system prompt or tooling made you make this mistake and cloud will kind of like go introspective”

Zevy Arnovitz · 46:30
#post-mortem#prompting#documentation#iteration