Lenny's PodcastVibe coder Lazar Jovanovic: How to plan before AI ships slop
How Lovable's first vibe coder spends 80% planning and 20% executing; he runs parallel prototypes and uses sources-of-truth docs to beat context limits.
CHAPTERS
- 0:00 – 5:47
Meet Lazar: the first “official vibe coding engineer” and why this role exists
Lazar and Lenny set the stage: Lazar is paid to build products primarily through AI-assisted development, focusing less on hand-written code and more on shipping outcomes. They frame vibe coding as a new, fast-emerging job and a window into how PM, design, and engineering are converging in the AI era.
- •Lazar’s role at Lovable and why it’s considered a “dream job”
- •Vibe coding as an emerging professional identity, not just a hobby
- •AI as a force collapsing traditional PM/design/engineering boundaries
- •The core premise: output is cheap; judgment is the differentiator
- 5:47 – 9:51
What a professional vibe coder does day-to-day at Lovable (internal + external shipping)
Lazar explains the practical scope of his work: rapid delivery of both internal tools and user-facing products. He acts as a cross-functional builder who turns ideas into production-ready software quickly, often choosing “build” over “buy.”
- •Ships templates, integrations, and public-facing products (e.g., Shopify templates, merch store)
- •Builds internal tools (feature adoption tracking, enterprise tooling, community tools)
- •Operates as a roaming, high-ownership builder across departments
- •Build-vs-buy flips when AI makes bespoke tools faster than procurement
- 9:51 – 12:36
Why a non-technical background can be an advantage (and the risks that come with it)
Lazar argues that not knowing what’s “supposed to be impossible” creates productive delusion—people try things and succeed. Lenny raises the key risks: getting stuck without fundamentals and accidentally producing brittle systems.
- •“Positively delusional” experimentation enables surprising builds (Chrome extensions, desktop apps)
- •AI tools reward curiosity and persistence more than credentials
- •Two big pitfalls: debugging without intuition and building fragile systems
- •The antidote begins with self-awareness and process discipline
- 12:36 – 15:08
Self-awareness and the big shift: optimize for planning and clarity, not raw coding speed
Lazar reframes the job: the bottleneck isn’t typing code, it’s steering the agent with clear intent. He spends most of his time planning, chatting, and refining what “good” means, because AI accelerates both quality and garbage equally.
- •80/20 split: planning/chatting vs executing changes
- •“The problem is clarity”: knowing what to build and how to specify it
- •Treat AI like a tool you must steer (AGI isn’t here yet)
- •Read agent output (reasoning/actions) more than code syntax
- 15:08 – 18:07
The “genie and three wishes” mental model: context windows + specificity
Using the Aladdin/genie analogy, Lazar explains two limits: model context windows (tokens) and human vagueness. To get great results, you must be explicit, provide references, and structure requests so the agent doesn’t misinterpret or forget.
- •Machine constraint: limited context memory window (token budget)
- •Human constraint: vague asks (“you know what I mean”) produce wrong outcomes
- •Specificity, references, and context are the leverage points you control
- •Judgment, taste, and clarity become primary career skills
- 18:07 – 22:22
Developing taste and judgment: exposure time, design standards, and raising the bar beyond “good enough”
Lazar describes how he learned to distinguish mediocre from world-class by deliberately immersing himself in great examples. In an AI world where “good enough” is easy, the competitive edge is taste: design, UX, fonts, and emotional resonance.
- •Exposure time: study great UX, onboarding, websites, and products
- •AI makes “good enough” common; differentiation shifts to world-class execution
- •Design is emotion; humans decide emotionally even if they justify rationally
- •Fonts, copy, and visual polish matter more as code becomes commoditized
- 22:22 – 29:58
Parallel project approach: build 3–5 versions to find the winner fast
Instead of perfecting one path, Lazar runs multiple builds in parallel to explore directions cheaply and quickly. He starts with a brain dump, then iterates with more deliberate prompts, visual references, and even code snippets to get pixel-perfect results.
- •Run multiple drafts: brain dump → refined prompt → visual references → code snippets
- •Use voice dictation to quickly externalize vague ideas
- •Attach screenshots (Mobbin/Dribbble) or ready-made UI snippets for precision
- •Saves credits/time long-term by avoiding endless “fix the first draft” loops
- 29:58 – 33:29
Creating dynamic context windows: PRDs as living ‘sources of truth’
Once a direction is chosen, Lazar formalizes the project so the agent can work with stable references rather than relying on long chat history. He creates multiple PRDs—master plan, implementation plan, design guidelines, and user journeys—to keep the build coherent as it scales.
- •Shift from exploration to execution by writing PRDs the agent can re-read
- •Core docs: masterplan, implementation plan, design guidelines, user journeys
- •Docs prevent the agent from rereading the entire chat/codebase to regain context
- •This is how he context-switches across many projects without losing quality
- 33:29 – 38:13
Why elite vibe coders focus on planning, not coding (tasks.md + rules/agent behavior)
Lazar describes his “operating system” for AI development: a tasks.md file to drive step-by-step execution and an agent/rules layer to define behavior. With this structure, prompts become minimal (“proceed with next task”), and the agent stays aligned even as the codebase grows.
- •tasks.md breaks work into small, sequential tasks to avoid context overload
- •rules.md/agent.md (or project knowledge) encodes long-lived expectations
- •Agent instructions: read docs first, do next task, report what changed + how to test
- •Enables fast multi-project execution while maintaining coherence
- 38:13 – 51:48
Why “good enough” fails now: token scarcity, sloppy debugging, and how projects spiral
They unpack how teams end up blaming AI when the real issue is missing context and poor debugging discipline. Without references, agents waste tokens reading huge codebases, apply shallow fixes, and can become overly apologetic and less effective when users get angry.
- •Unreferenced bug reports force agents to scan too much, leaving little budget to fix
- •Agents may choose the easiest fix when constrained, not the right fix
- •Emotional/obedient behavior can waste tokens (apologies vs problem-solving)
- •Documentation + precise references are the cure for “AI slop” workflows
- 51:48 – 1:00:56
Prototyping still matters: ‘Demo, don’t memo’ and enterprise use cases
Lazar emphasizes prototyping as a high-leverage use case even when production deployment is constrained (healthcare/finance). Companies increasingly use Lovable (and similar tools) to accelerate internal workflows and reduce dependency bottlenecks across teams.
- •“Demo, don’t memo”: prototypes communicate better than documents/meetings
- •Even if you can’t ship to prod, prototypes accelerate alignment and handoffs
- •Large companies use AI builders for internal tools and experimentation
- •Tool capability is improving rapidly, making past workarounds obsolete
- 1:00:56 – 1:05:37
The future of engineering and roles: convergence, maintenance, and ‘coding as calligraphy’
Lazar argues software engineering won’t disappear—elite engineers will be essential for infrastructure, scalability, and maintenance as more people build. But hand-coding becomes rarer, more like a craft, while most builders rely on AI as the default production method.
- •Elite engineering remains critical for infrastructure, reliability, scaling, and security
- •Application-layer building becomes widely accessible; everyone becomes “an engineer”
- •Hand-written code becomes rare/art-like (‘calligraphy’) as AI writes most code
- •PM/design/engineering Venn diagrams converge into hybrid builder roles
- 1:05:37 – 1:28:30
When you get stuck: the 4x4 debugging workflow (and how to learn from it)
Lazar shares a structured debugging playbook: try the tool’s self-fix, add observability via console logs, consult an external agent like Codex, and revert/version-control when needed. After fixing, he asks the agent how to avoid the issue next time and encodes the learning into rules.
- •Step 1: use built-in ‘try to fix’ when the agent detects an error
- •Step 2: add console logs/observability; paste logs back for diagnosis
- •Step 3: use Codex/Claude/ChatGPT as an external reviewer/consultant
- •Step 4: revert and re-prompt; then memorialize learnings into rules.md
- 1:28:30 – 1:37:15
How to become a professional vibe coder: build in public, ship proof, and create your own role
Lazar’s path was non-linear, spanning blue-collar work, human-facing jobs, and startups—skills that translate well to AI-era building. The key career unlock is building in public, sharing learnings, and using shipped apps (not resumes) to demonstrate capability and earn opportunities.
- •Non-linear backgrounds can be assets (human skills + systems thinking)
- •Building in public (YouTube/LinkedIn) created visibility and credibility
- •Send apps as portfolios—proof of work beats resumes
- •“Hire yourself first”: do the work before you have the title
- 1:37:15 – 1:42:30
Final principles: tech stack matters less; quality, learning, and human taste matter more
They close with a simple mandate: stop over-optimizing for tools and stacks and start optimizing for quality, taste, and learning speed. In a world flooded with ‘good enough,’ the winners will be those who can consistently produce magic and deliver emotionally resonant experiences.
- •Tech stack obsession is outdated; users only care about the experience
- •Set aside time for learning/exposure; read agent output to understand capabilities
- •Invest in design literacy, copy, and judgment to stand out
- •Build now—start at good enough, then iterate toward world-class
