Lovable

Lovable hired a non-coder to ship products across the company using AI alone. Lazar Jovanovic, Lovable's first official vibe coding engineer, has never written code beyond a couple of console logs. He built most of the Shopify integration templates users remixed, Lovable's merch store, and an internal feature adoption matrix tracking how many people use each new feature.

He started in growth working for Elena Verna, who had many ideas and needed someone to own them, and is moving into go-to-market, enterprise and community tools. When setting up an off-the-shelf enterprise tool would take an hour or two, he builds his own version. He says Lovable is growing so fast that "every department needs a Lazar now or yesterday" and thinks at least half of S&P 500 companies have people using it.

Lenny's Podcast

Hand off ideas as working prototypes instead of memos. Lovable's 2025 motto was "demo don't memo." Lazar built a prototype in four hours that engineers replicated in production six to seven months later; he estimates writing it up would have taken a week or two.

He says prototyping is useful even when regulation blocks production use, because it replaces documents and meetings as the way to communicate an idea.

Lenny's Podcast▶ 53:42

Not knowing what is impossible lets non-engineers build surprising things. Non-technical users built a Chrome extension and desktop apps in Lovable while engineers explained why it shouldn't work, and Lovable's community manager prompted her way to generated video before it was a feature.

Lazar calls the mindset "positively delusional": everything is possible until proven wrong. He pairs it with self-awareness about your gaps and checking feasibility before brute-forcing, since the risks are getting blocked with no fix or shipping a fragile build you can't judge.

Lenny's Podcast▶ 1:03:18

Lovable hires people who build in public and send apps, not resumes. Elena Verna picked Lazar out of many vibe coders because he shared projects, failures and knowledge on YouTube and LinkedIn and was doing the job before being hired. Several other hires stood out by sending Lovable apps instead of resumes.

Lovable employees always open a lovable.app link. Quote: "You don't need a company to hire you. You can hire yourself as a professional vibe coder first."

Lenny's Podcast▶ 1:06:19

Spend 80% of the time planning and chatting, 20% executing. Lazar uses chat mode to clarify what he wants before any code is written, then lets the agent execute the plan. He says the problem to solve is clarity, not coding, since AI output is already faster than human output.

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

Lenny's Podcast▶ 27:50

Agents fail from small context windows and literal reading of vague wishes. Lazar's genie model: each request has a limited token budget spent on reading, browsing, thinking and executing, and a vague wish like "make me taller" gets you a 13-foot man. In a large codebase with no references, the agent can burn about 80% of its tokens just reading, then take the easiest fix.

You can't control the context window or model quality, only the human side: being specific, giving references and providing context. He admits his token-budget figures, such as an arbitrary 100,000 tokens, are guesses.

Lenny's Podcast

Build four or five prototypes in parallel, then keep the clear winner. He starts projects at once from a voice brain dump, a typed and thought-through prompt, a prompt with a reference screenshot from Mobbin or Dribbble, and a prompt with a code snippet from a library like 21st.dev. He says it costs a few credits up front but saves hundreds of credits and days over polishing a weak first build.

Code snippets steer the AI toward more pixel-perfect results than describing a design in words, and comparing several designs avoids generic AI slop. He began it because he was waiting on agents anyway, and says everyone he tested it with called it an eye opener.

Lenny's Podcast▶ 29:23

Planning documents carry context so the prompt becomes "proceed with the next task". After picking a winner, he has an LLM write a master plan, an implementation plan, design guidelines with some CSS specifics, and a user journey, then a tasks.md with tasks and subtasks. Project rules tell the agent to read the docs, do the next task, and report what it did and how to test it, letting him run five or six projects at once.

He does this because by message 10 to 40 early instructions get lost as the agent optimizes for speed over rereading. He calls the documents "sources of truth."

Lenny's Podcast▶ 33:30

When everyone can build good enough, taste sets builders apart. Lazar says the gap between good enough and world class has widened because everyone now produces good enough, and that design and emotion now matter more than code. He builds judgment through "exposure time" with content and people that raise his standards.

His first weeks went from "I can build" to "I can build fast" to "should I even have built this?" Caveat: the workflow rests on one Lovable employee's experience with no comparative data, and it skips the security and maintenance risks of shipping code nobody reads.

Lenny's Podcast

Lovable starts coding immediately, so force a planning phase on complex work. Zevi Arnovitz says Bolt and Lovable have coding-agent system prompts that write code right away, which works early but caused gnarly bugs in his apps on payments and database changes. He has the AI read the codebase and ask clarifying questions about scope, data model and UX before writing a plan.

Quote: "planning is really important when you're implementing something technical."

Lenny's Podcast▶ 27:30

Lovable's harness trades control for ease of building. Bolt, Lovable, Replit, Base44 and v0 use the same underlying models but add layers that decide things like sign-in and databases for you. Cursor or Claude Code give full tool access and the latest model abilities, at the cost of making those decisions yourself.

Arnovitz suggests graduating from a chatbot to Bolt or Lovable to Cursor to the terminal, moving up when you outgrow a tool; he moved off Bolt when connecting payments became too much. Quote: "the main difference between all these tools is basically the harness."

Lenny's Podcast▶ 1:16:10

Lovable is one of many coding agents that work, not a lone winner. Anish Acharya names Claude Code, Codex, Lovable and Replit as coding agents that are all working, as evidence that each layer of the stack has about 20 relevant players.

Lenny's Podcast▶ 1:39:05

A free year of Lovable in a bundle drew fraud rings. Lenny's Product Pass launched with a free year of Cursor, Lovable, Bolt, Replit and v0, which he calls way too good an offer. Fraud rings, mostly in China, found API exploits and spread them through student networks, and he had to work with Stripe and Substack to shut them down.

His engineer barely slept for a week patching holes, and Lenny calls it the most stressful business moment he has had. Quote: "that was way too good an offer. People came for it, bad people."

Lenny's Podcast▶ 52:45