How a non-technical vibe coder landed a full-time AI-building job

Executive overview

Most people using AI tools optimise for raw output speed. The real constraint is clarity — being specific enough that the model builds the right thing, not just any thing quickly.

Lazar Jovanovic, professional vibe coder at Lovable, has never written code. His edge is spending 80% of his time planning and only 20% executing. Non-technical builders are often better at this because they approach AI without assumptions about what is or isn't possible.

The ceiling on AI output isn't model intelligence — it's what the model sees before it acts.

The parallel-build method for rapid clarity

• Start with a voice brain-dump prompt in a new project — don't refine, just send.
• Open a second project with a typed, more deliberate prompt once threads become clear.
• In a third project, attach a screenshot from Mobbin or Dribbble to show the target design quality.
• In a fourth, find an existing code snippet or template and attach it directly — models interpret code more precisely than English descriptions.
• Compare all four; the winner is usually obvious within a couple of prompts.
• This approach costs more credits upfront but saves hundreds of credits by avoiding weeks of course-correction.

Context management and the genie problem

• LLMs have a finite token window — every token spent reading, thinking, and apologising is a token not spent solving.
• Vague prompts force the model to spend ~80% of its window just getting oriented, leaving little capacity for the actual fix.
• Insulting the model when it fails wastes tokens on anxiety management rather than problem-solving.
• The solution is to externalise context into files the agent reads every session, not to rely on in-conversation memory.

The PRD and rules file system

• Master plan (.md) — high-level intent: what is being built, for whom, and how it should feel.
• Implementation plan (.md) — the order of operations: backend first, then auth, then APIs, etc.
• Design guidelines (.md) — includes CSS parameters and named design styles (glassmorphism, Bauhaus) to constrain the model's creativity.
• User journeys (.md) — step-by-step flows from registration through key actions.
• Tasks.md — the single source of truth for what to do next; derived from all of the above.
• Rules / agent.md — tells the model how to behave every session: read all PRDs first, check tasks.md for the next task, report what was done and how it was tested.
• Once these files are in place, prompts shrink to "proceed with the next task" — context is delegated to the agent, not carried in conversation.

The four-by-four debugging framework

• Step 1 — Let the tool self-correct. Most tools surface an error state with a one-click fix; use it first.
• Step 2 — Add console logs. Ask the model to instrument the relevant files, re-run the broken function, then paste the log output back into chat.
• Step 3 — Use an external model. Export the codebase to GitHub, import into Codex or compress with Repomix and upload to Claude or ChatGPT for an independent diagnostic — do not let the external tool make edits, only diagnose.
• Step 4 — Revert and re-prompt. Use built-in version control to go back three steps, then rewrite the prompt more clearly after a break.
• After any fix, ask the model: "How could I have prompted you better to solve this in one step?" Then add the answer to rules.md so the agent learns it permanently.

Skills that will hold value as AI takes on more

• Judgment and taste — knowing what world-class looks like, not just good enough; the gap between good enough and world-class has collapsed because everyone produces good enough with AI.
• Design literacy — exposure to exquisite design, learning named styles, understanding that a "simple" gradient may be 50 layers of opacity; follow elite designers and watch them work.
• Copywriting (human voice) — AI-written copy is already detectable within a few sentences; authentic human voice is increasingly scarce.
• Emotional intelligence — human-to-human interactions, live experiences, and comedy are areas AI cannot replicate; anything deterministic (translation, middleware, rote journalism) is at risk.
• Building in public — distribution is as scarce as attention; sharing failures and frameworks openly is what creates career opportunities in this space.

The professional vibe coder career path

• Build in public on LinkedIn, YouTube, or wherever your format fits — document the failures, not just the wins.
• Participate in hackathons and local builder communities to create serendipity.
• Apply to companies by sending a Lovable app that demonstrates fit, not a resume.
• Do the job before being hired — Lazar was already shipping production-grade tools before Lovable formalised the role.
• Tech stack choices are irrelevant; end users want a stellar experience, not a specific framework.