Vibe Coding: Why Product Managers Should Rethink Their Team Setup

His engineers asked him to teach them how he works.

Not the other way around.

Zevi Arnovitz is a product manager at Meta. Not a developer. No technical background. Yet he plans, builds, reviews, and ships complete features and products – without writing a single line of code himself.

His approach: he treats AI like his CTO. He owns the problem, the user experience, and the product decisions. AI owns the technical approach and implementation.

The result was so compelling that his engineers – the people who actually know how to code – asked him to walk them through his workflow.

He's not an outlier. Products built by non-developers using AI are popping up everywhere. Complete web apps. Backend, frontend, payments, user management. Not a single line written by hand. At Squills, ai-portrait.eu – a full AI portrait app – was a first successful test: built entirely with vibe coding, without traditional development resources.

The question is no longer whether this works. The question is: what does it mean for everyone working in product teams?

What is vibe coding – and what is it not?

The term sounds like a trend, like hype, like the next buzzword that'll be forgotten in three months.

It's not.

Vibe coding describes a fundamentally different approach to software development: you describe in natural language what you want to build. AI tools like Cursor, Claude, or GitHub Copilot generate the code. You steer, you decide, you set the direction.

Sounds like "AI does everything for you"? That's exactly what it's not.

What vibe coding really is: a new form of collaboration between human and machine. You no longer need a computer science degree. But you need something else – something most people underestimate:

The ability to articulate problems with crystal clarity.

Because AI writes exactly the code you describe. Not the code you mean. Not the code you have in your head. The code you put into words. And that's where the wheat separates from the chaff.

Arnovitz demonstrates this: for every exploration phase, he prepares thoughtful answers – how scoring should work, what percentage uses which question type, what the UX should look like. The more precise the input, the better the output. That's not development work. That's product work.

The uncomfortable question for product teams

If non-developers can build functional apps – what does that mean for the composition of your team?

Developers aren't becoming obsolete. That would be naive and wrong. But the line between "can code" and "can't code" is disappearing. And that changes which competencies form the bottleneck in a product team.

Before: You had an idea. You needed a dev team to build it. The bottleneck was development capacity.

Now: You have an idea. You can prototype it in hours. The bottleneck is no longer execution. The bottleneck is judgment.

Which problem is worth solving? Which feature delivers real user value? When is a product "good enough" to ship? These questions matter more than ever. Because execution has become so cheap that you can afford mistakes faster – but you also fall on your face faster.

The numbers behind it

This isn't gut feeling. The data is clear.

According to Carta, the share of solo foundings among all new startups grew from 23.7% in 2019 to 36.3% in the first half of 2025. Every third new startup today is founded by a single person.

AI startups reach one million dollars in annual revenue up to four months faster than traditional SaaS companies. Infrastructure costs that consumed thousands of euros per month just five years ago now often sit below 100 euros.

A concrete example: Maor Shlomo founded Base44 in early 2025. An AI-powered no-code app builder. As a one-person startup. Six months later: 300,000 users, 3.5 million dollars in annual revenue, acquired by Wix for 80 million dollars. Starting capital: roughly 10,000 dollars.

Sam Altman has a bet going among tech CEOs about when the first one-person billion-dollar company will emerge. Sounds like Hollywood.

But the really interesting story isn't the unicorns. It's the thousands of individuals and tiny teams currently building real, profitable products with AI. Six-figure, seven-figure revenues. Bootstrapped. Profitable. Sustainable.

If a single person can do that – what does it mean for a team of twenty?

How vibe coding works in practice

Just opening ChatGPT and saying "build me an app" doesn't work. What works is a system.

The workflow: AI as CTO

Arnovitz has broken his entire development process into phases and turned each phase into reusable prompts:

1. Create issue – Capture ideas quickly while working on something else. AI automatically creates a structured ticket.
2. Exploration – AI analyzes the codebase, explains the current state, and asks targeted questions about scope, UX, and architecture.
3. Create plan – A detailed markdown plan with critical decisions and a task list that gets updated during execution.
4. Execution – Different AI models for different strengths: Claude for architecture and reasoning, other models for fast coding or UI work.
5. Review and peer review – Multiple AI models review the code independently. A multi-model review panel where AI agents debate bugs and design decisions.
6. Learning and documentation – Every mistake becomes a learning opportunity. Every architectural decision gets documented.

The result: StudyMate – an app that lets students automatically generate challenging quizzes from PDFs and notes. Real users, real revenue. Built by a PM who doesn't write code.

This principle generalizes. A multi-agent setup where different AI roles handle different tasks:

• Architect: Plans the system architecture. Data model, API structure, tech stack decisions.
• Developer: Writes the code based on architecture specifications.
• Reviewer: Checks the code for bugs, security vulnerabilities, and consistency.
• Tester: Identifies edge cases and writes tests.

Sounds like a normal development team? Exactly. The difference: all of these roles are handled by AI. The human is the product manager who sets the direction.

The rules: no structure, no results

Regardless of setup – without clear rules, vibe coding produces chaos. AI is incredibly fast. But without guardrails, it builds a pile of technical debt just as fast.

What works:

• Clear architecture documents before the first prompt. AI needs context. A lot of context.
• Small, self-contained tasks instead of "build me the entire feature." The more precise the prompt, the better the result.
• Code reviews by a separate AI instance. Never let the same agent write and review code. That's like having the author proofread their own book.
• Version control from day one. Git, commits, branches. Without rollback capability, you're lost when AI heads in the wrong direction.
• AI-native postmortems. When AI makes a mistake: what in your context led you to this error? Then adjust prompts, documentation, and workflow. That's how the system improves over time.

The result: weeks instead of months

Arnovitz's StudyMate is live. Users upload PDFs, generate quizzes, learn with them. At Squills, ai-portrait.eu is up and running – users sign up, upload photos, generate AI portraits, pay. Not perfect. But functional, usable, and live.

Three years ago, projects like these would have kept a team of three to five developers busy for several months. Today: one person, a few weeks, and the product is on the market.

What this means for the job market

And now the question nobody likes to ask.

If one person with AI can do the work of a small team – what happens to the people who've been doing that work?

The comfortable answer "new jobs will emerge" is partially true. But it's also an excuse.

The honest answer: certain skills are rapidly losing value. Pure coding – translating clear specifications into working code – is the task AI does best. Anyone who defines themselves exclusively through this skill has a problem.

At the same time, other skills are rising in value:

• Problem understanding: Knowing what should be built. User interviews. Market analysis. Product discovery.
• Systems thinking: Understanding how components interact. Architecture decisions. Weighing trade-offs.
• Quality judgment: Recognizing when something is "good enough" and when it's not. Code review. UX sensitivity.
• Communication: The ability to articulate requirements so that an AI (or a human) can execute them. That's harder than it sounds.

Arnovitz puts it this way: PMs should be "10x learners," not "10x geniuses." AI helps practice strategy, UX, and technical collaboration at a much higher level and pace. The ability to learn becomes more important than existing knowledge.

For product managers, this means: your core competency has never been more valuable than right now. But your role is changing. You're no longer the translators between business and engineering. You're the conductors of an orchestra where AI agents increasingly play the instruments.

What doesn't work

Before anyone gets the impression that tomorrow everyone with an AI subscription will build an app: no.

Vibe coding has clear limits. And anyone who doesn't know them wastes time.

Complex, safety-critical software: Banks, medical devices, flight control systems. These still require deep technical expertise and strict certification processes. AI can assist, but not lead. Heavy database migrations and high-risk changes in large organizations still belong under rigorous engineering review.

Highly scaled systems: When millions of users access a system simultaneously, "works on my machine" isn't enough. Performance optimization, caching strategies, infrastructure design – these are disciplines that require experience and deep technical understanding.

Anything without a clear problem: The best tool in the world doesn't help if it's unclear what should be built. Vibe coding accelerates execution. Problem discovery remains human work.

And one more thing: the quality of the output depends directly on the quality of the input. Vague prompts produce vague code. Anyone who doesn't understand what AI generated can't debug it when things go wrong. "I don't understand code" isn't a free pass. It's an invitation to learn at least the basics.

What product managers should do now

Concrete recommendations. No theory, no vision – things you can start this week.

1. Build a prototype yourself

Not to replace your developers. But to understand what's possible. Pick a small problem that annoys you. An internal workflow, a simple tool. And build it with Cursor or a comparable tool.

Two things will become clear: first, how much faster execution has become. Second, how much harder it is to describe a problem cleanly than you thought.

2. Question your team setup

Not to fire people. But to ask: if simple coding can be handled by AI – is development capacity being invested in the right tasks? Are your best engineers working on problems that truly require their expertise? Or are they writing CRUD endpoints that AI generates in seconds?

3. Invest in architecture competency

What matters most in vibe coding isn't the coding itself. It's the architecture. Clean system designs, clear interfaces, well-thought-out data models. These are the specifications that determine whether AI produces usable or catastrophic code.

The most valuable competency on your team will soon no longer be "can code fast," but "can think clearly."

4. Build an AI-native workflow

The key isn't the individual tool – it's the workflow. Arnovitz spent months refining his process – from issue creation through exploration and planning to multi-model reviews. He turned this workflow into reusable prompts that structure every phase.

That means: learn to write good prompts. Learn to read AI-generated code. Understand when to trust AI and when not to. And above all: improve the workflow with every mistake.

The shift that's happening right now

AI isn't replacing developers. AI is changing the role of developers. Away from pure code production, toward architecture, system design, and quality assurance.

And product managers play a key role in this new world. Because the ability to identify the right problem, prioritize the right solution, and determine the right quality level – that's exactly the bottleneck AI doesn't solve.

The most exciting development in digital product development isn't that AI makes everything easier. It's that AI shifts the bottleneck. Away from execution. Toward judgment, taste, and clarity of thought.

These are skills you can learn. But you have to start.

This article appears on the Squills blog. Squills builds AI-powered products and helps teams work faster with AI. Want to know how vibe coding can work in your team? Talk to us.