All podcasts / Lenny's Podcast / Summary

The rise of the professional vibe coder (a new AI-era job) | Lazar Jovanovic (Professional Vibe Coder)

2026-02-08 - 102 min - source - Read full transcript
Lenny Rachitsky (host)Lazar Jovanovic

Key insights

A non-technical background can be an advantage for AI-assisted building because you don't carry assumptions about what 'shouldn't' be possible.
Lazar argues people without engineering backgrounds get further because they simply try things (Chrome extensions, desktop apps, in-app video generation) that technical people rule out in advance as impossible with a given stack. He calls this being 'positively delusional' - assuming everything is possible until proven wrong.
vibe-coding-workflow
Before committing to a direction, run several parallel prototype starts from the same idea rather than iterating on one.
Lazar's method: start one project from a raw voice brain-dump, a second from a more deliberate typed prompt, a third seeded with a design reference from Mobbin or Dribbble, and a fourth from an actual code snippet template. Comparing the outputs converges on clarity faster and cheaper than refining a single early attempt, because AI code and design still communicate best in code, not English.
vibe-coding-workflow
The AI's limited context window works like the three-wishes rule in the Aladdin story - vague asks produce badly-scoped results, not because the model is weak but because the ask wasn't specific.
Lazar's genie analogy: ask for 'taller' and the genie makes you 13 feet tall - technically correct, functionally useless. He treats token budget as the scarce resource: every request splits a fixed allocation between reading context, browsing, thinking, and executing, so vague prompts waste most of it on disambiguation rather than output quality.
ai-context-and-clarity
A layered PRD system (master plan, implementation plan, design guidelines, user journey, tasks.md) lets an operator delegate ongoing context management to the agent itself, enabling several projects to run in parallel.
Once these documents exist and the agent is instructed to read them before acting, Lazar's own prompts shrink to 'proceed with the next task' - he stops being the context-holder and just reads agent output, which is how he runs up to six Lovable projects simultaneously without losing quality.
ai-context-and-clarity
When debugging, escalate through a fixed four-step sequence before assuming the tool is broken: try the built-in fix, add your own console logging for awareness, bring in an external diagnostic tool, then revert and re-prompt more clearly.
Lazar's 'four by four': (1) click the agent's own suggested fix once, (2) if the agent is unaware anything is wrong, manually add console logging to surface the failure, (3) if that's still not enough, export the code to GitHub and hand it with logs to an external reviewer like Codex or paste a RepoMix-compressed codebase into ChatGPT/Claude for diagnosis only, (4) if none of that works, assume the prompt itself was the problem, revert, and try again with a clearer ask. He estimates 90-100% of persistent bugs trace back to an unclear or under-referenced prompt, not a tool limitation.
ai-context-and-clarity
After a bug is fixed, ask the agent to translate the fix into a permanent rule so the same failure mode doesn't recur - the goal is to remove yourself from having to remember anything.
Lazar asks the agent, 'how could I have prompted you better so we solve this in one go next time,' then has it write that answer directly into rules.md. Because the agent reads rules.md on every task anyway, the lesson persists automatically instead of depending on the human remembering it two weeks later.
ai-context-and-clarity
AI has collapsed the cost of 'good enough' output to near zero, which means the competitive gap has moved entirely to taste, design, and judgment.
Lazar's framing: a decade ago, simply shipping something working was rare and valuable. Now everyone using AI tools can produce 'good enough' by default, so the meaningful gap is no longer between bad and good, but between good-enough and world-class - and closing that gap requires human taste the model can't supply on its own.
design-taste
Design and copywriting taste is best built through deliberate 'exposure time' - consciously consuming great examples - not just through building.
Lazar credits Vercel's Guillermo Rauch with the concept. He personally follows designers who publish their prompting and design process, studies real component libraries at the layer level (finding a single 'simple' gradient was actually 50 layered colors), and built his own app just to teach himself named design styles like Bauhaus and glassmorphism.
design-taste
Hand-writing code will become a rare, admired craft - like calligraphy - rather than the core skill that determines who can build software.
Lazar predicts that within roughly a year 'vibe coding is just coding,' and that most elite engineers already rely on AI for the majority of code output. Writing code by hand will persist as a niche, almost artisanal skill rather than the primary lever for shipping products.
future-of-tech-roles
Elite, foundational software engineering remains essential even as building becomes accessible to everyone, because someone still has to build and maintain the infrastructure that supports mass building.
Lazar points to Cloudflare outages and Lovable's own scaling pains as evidence that hosting, security, APIs, and core infrastructure still require deep engineering skill, and that this need grows, not shrinks, as more non-engineers start shipping products on top of that infrastructure.
future-of-tech-roles
The historically separate PM, engineering, and design functions are converging into a single AI-assisted builder role centered on judgment rather than specialized execution.
Lazar argues everyone becomes a kind of 'forward deployed engineer' regardless of title, because AI now handles the execution layer that used to differentiate the three roles. He bets designers become the next big winners after PMs, since AI is improving at technical decisions faster than it is at emotional/aesthetic ones, and design is fundamentally about emotion.
future-of-tech-roles
You don't need a company to hire you into a vibe-coding role - build in public and demonstrate the skill directly, including by submitting a working app instead of a resume.
Lazar got hired at Lovable by already doing the job publicly: a YouTube channel documenting failures and projects, heavy LinkedIn posting, and sharing his workflow openly rather than treating it as a secret. He notes Lovable has hired candidates who sent a built Lovable app in place of a resume, and that his own advice to aspiring vibe coders is simply to 'do the job you would have done anyway.'
vibe-coding-workflow

Media referenced

Companies

Techniques and frameworks

Summary

Lazar Jovanovic is, by his own description, the first person hired at Lovable with the explicit title of vibe coding engineer - a self-taught, non-technical builder who spends his full-time job shipping internal and external products entirely through AI coding tools like Lovable, having never written a line of code by hand. The conversation opens by dismantling a common assumption: that a technical background is an advantage when working with AI coding tools. Lazar argues the opposite - not knowing what "shouldn't" be possible let him build things (Chrome extensions, desktop apps, in-app video generation on Lovable before it was a supported feature) that technically-minded people ruled out in advance. He calls this staying "positively delusional," paired with enough self-awareness to know the real bottleneck isn't code generation but clarity of the ask.

Much of the episode is a practical workflow breakdown. Lazar starts new projects by running several parallel prototypes from the same idea - a raw voice brain-dump, a more deliberate typed prompt, one seeded with a design reference pulled from Mobbin or Dribbble, and one seeded with an actual code snippet - because AI tools interpret code and visual references far more reliably than natural language alone. Once a direction is clear, he shifts into what he calls the planning phase: building a stack of markdown documents (a master plan, an implementation plan, design guidelines, a user-journey doc, and a granular tasks.md) plus standing rules.md/agent.md instructions that tell the agent how to behave without re-explaining context every time. This is what lets him run five or six Lovable projects simultaneously - once the documentation exists, his own prompts shrink to "proceed with the next task," and he spends his attention reading agent output rather than writing code or micromanaging execution.

He frames the AI context window through a genie-and-Aladdin metaphor: ask vaguely and you get a technically-correct but useless result, the way a genie who grants "taller" makes you thirteen feet tall. Token budget is treated as the truly scarce resource, split between reading references, thinking, and executing, so unclear prompts waste most of it on disambiguation. When something breaks, he uses a fixed four-step escalation - the tool's own fix button, manually added console logging to build the agent's awareness, an external diagnostic pass through OpenAI's Codex or a RepoMix-compressed codebase pasted into ChatGPT or Claude, and finally reverting and re-prompting on the assumption that most persistent bugs trace back to his own unclear ask. After any fix, he has the agent write the lesson directly into rules.md so it never has to be re-taught.

The back half of the conversation turns to where this leaves human skill. Lazar's central claim is that AI has collapsed the cost of "good enough" output to near zero, so the competitive gap has shifted entirely from functional-versus-broken to good-versus-world-class, and closing that gap requires taste the model doesn't supply on its own. He credits Guillermo Rauch's concept of "exposure time" - deliberately consuming great design and copy rather than only building - as the practice that develops judgment, and describes discovering that a "simple" gradient he admired was actually fifty layered colors, a moment that reset his sense of what quality actually requires. He predicts hand-written code becomes a rare, admired craft like calligraphy, that elite engineering remains essential for the infrastructure everyone else builds on top of, and that PM, engineering, and design converge into a single AI-assisted builder role - with designers, not engineers, likely the next big winners, since AI is improving faster at technical decisions than emotional ones.

On career path, Lazar's advice is to build in public rather than wait to be hired: his own route into Lovable came from documenting failures and workflows on YouTube and LinkedIn before the company approached him, and he notes candidates have gotten hired by sending a working Lovable app in place of a resume. His closing framing is unusually optimistic relative to the anxiety many builders express about AI: he sees the shift as finally letting people get paid for clarity and judgment - the parts of work most people actually enjoy - rather than for repetitive execution.

Notable Quotes

"Coding is going to be like calligraphy. It's going to be so rare that it's going to become an art." - Lazar Jovanovic

"We won't be rewarded in the world of AI for faster raw output. We will be rewarded for better judgment." - Lazar Jovanovic

"If a tool is super powerful and you give it a wrong input, the output's going to suck as well." - Lazar Jovanovic

"It feels like an emerging core skill is learning clarity in the ask of the AI." - Lenny Rachitsky

"AI is never going to write good comedy. It's impossible. It just doesn't have that layer." - Lazar Jovanovic