Pat Walls spent 8 years building Starter Story.
Thousands of hours. Tens of thousands of lines of code. Over 7,000 commits.
Then a guy who has never written a single line of code sat down next to him and said he could clone the whole thing in under 60 minutes.
Not a landing page. The actual product: case studies, an ideas library, a full academy, a community, auth, memberships. Pat says the academy feature alone took hundreds of hours, maybe a thousand.
The guy's name is Lazar. He works ops at a startup during the day. At night he vibe codes, and he's shipped over 50 projects in the last 6 months without writing code once.
The stunt is fun. The reason I'm writing about it is the vibe coding workflow Lazar runs, step by step, on camera. It's the most transferable system I've seen for building with AI tools, and I'm going to break it down for you.
Watch the full experiment here:
The bet: 8 years of code vs 60 minutes of prompts
Starter Story is a custom web app built on Rails. Pat listed exactly what Lazar had to replicate:
- The custom design and look of the site
- Case studies (breakdowns of real businesses and how they grew)
- An ideas page with lists, examples, and validation resources
- The academy: a course where members submit milestones
- A community section
- Authentication: sign up, log in, password reset
Lazar pulled out a timer and started it on camera.
His plan sounded almost boring: start with the idea, break it into smaller pieces, write project documentation, build every page with mock data, then swap in a real database and auth at the end.
That boring plan is the whole trick. Most people open Lovable or Cursor, type "build me an app like X", and get slop. Lazar barely typed a creative prompt all session. He ran a system.
Step 1 of the vibe coding workflow: write 4 documents before you build anything
Lazar didn't start in a builder. He started in ChatGPT, with a "Lovable base prompt" GPT he found under Explore GPTs.
He fed it the concept: what he's building, the name, the target audience, the core features and pages. He even screenshotted the Starter Story homepage so the AI had a design reference. ("If I'm cloning it, let's clone it for real.")
Then he had it generate 4 documents:
- Master plan: what he's building. His analogy: I'm building a house.
- Implementation plan: how he's building it, in order. How am I building the house?
- Design guidelines: how am I going to paint the house?
- App flow, pages and roles: who does what. The engineer, the foundation guy, the landscaper.
First thing after Lovable spits out a project? He transfers it to GitHub, creates a docs/ folder, and commits all 4 documents into the repo.
His reason: "I want to own the code." And now every future prompt can just say "read the docs".
Pat's summary was perfect: it's like giving the project a brain.
Step 2: spend 80% of your time in chat mode, not build mode
This one surprised me the most.
Lovable has a default mode that writes code and a chat mode that just talks. Lazar says he spends about 80% of his time in chat mode.
Before letting the AI touch code, he makes it read the docs, confirm it understands the plan, and propose the first phase. Then he pushes back: revise the roadmap for my timeline, break it down into small steps, show me exactly how you're going to build the case study page.
Only when the plan looks right does he flip to default mode and say "proceed".
Pat called it "having a little meeting with it". That's exactly what it is. You're managing a very fast junior developer who lies about being ready.
And when the AI went rogue and built a "founders" page nobody asked for, Lazar didn't rage-prompt. He switched to chat mode and said: why did you build the founders page? Go back and read the master plan. The AI apologized and fixed its own roadmap.
→ The docs aren't bureaucracy. They're the thing you point to every time the AI drifts.
Step 3: have a separate debugging system ready before you need it
Around the middle of the build, Lovable shipped a bug it could not fix itself. This is the exact moment most beginners quit. Pat said it straight: he's seen tons of people give up as soon as the vibe coding gets complex.
Lazar has a two-part system for it.
Part one: a custom debugging GPT he built in ChatGPT. He doesn't even explain the problem. He dumps the raw error code in, and the GPT identifies the cause, proposes a fix, and then interrogates itself: "why do you think this will work?" Half the time it admits it's not sure and thinks again. The output is a ready-to-paste prompt for Lovable.
Part two, for when the AI is stuck in a loop: ask it to refactor. The error shows up again, but now it's in a smaller file. In this case the whole bug came down to one word on line 64. Changing "what's" to "what is" fixed the build.
Hours of potential rage, solved by shrinking the haystack.
Step 4: screenshots are the cheat code
Every time Lazar needed a page the AI couldn't guess, he didn't describe it. He screenshotted it.
The ideas page? Screenshot of starterstory.com/ideas, plus a screenshot of another site (Idea Picker) whose layout he likes. Then he labels them: image one is the category page, image two is the ideas within a category, image three is one idea broken down.
The academy, the feature Pat spent hundreds of hours on? Screenshots again, with a note: merge the best of both worlds with the plan you already have.
The memberships page? "I'm just going to rip you off completely here." Screenshot, paste, done.
And for polish, he pulled a component from 21st.dev, a library of prebuilt UI elements. Searched "global", grabbed an animated globe, copied the prompt, and dropped it into the community page. Pat's reaction on camera: "that looks so good."
→ A screenshot is worth roughly a thousand words of prompt. Stop describing layouts. Show them.
The backend took 2 clicks and one API key
Here's where I expected the experiment to fall apart. It didn't.
Lazar created a Supabase project (his database), went back to Lovable, and clicked connect. That's the entire backend setup. Two clicks.
Then he did something genuinely cheeky. He grabbed an OpenAI API key and had Lovable build an edge function: an AI assistant that takes any business idea on the site and generates a full plan for it. Market size, competitors, monetization, SWOT.
It worked. On camera.
Pat's reaction was the best moment of the video: "We don't even have a chatbot. After thousands of lines of code and thousands of hours, we don't have a chatbot."
The clone shipped a feature the original didn't have. A feature Pat says users had been requesting for a long time.
So did he actually clone a $1M business?
Honest answer: no. And yes. Depends what you mean.
They blew past the 60-minute mark, though Pat admits most of that was waiting for AI tools to finish generating. By the end, the clone had case studies, an ideas library, an academy, a community page with a fancy globe, a memberships page, a connected database, and a working AI assistant.
What it didn't have: Pat's 8 years of content, his SEO, his email list, his brand, his distribution. The features were cloned. The business wasn't. Nobody wakes up tomorrow and pays for the copy.
Pat knows this better than anyone, and he still wasn't dismissive. His take: "When I started my business, it was really just a copy of another business. I put my twist on it, built it, and got eyeballs and trust."
That's the real lesson, and it cuts both ways:
→ Your product features are no longer a moat. Anyone with Lazar's workflow can rebuild them in an afternoon.
→ Distribution and trust are the moat. Those still take years, which is exactly why they're worth building.
I keep seeing this pattern in founder stories. Kian vibe coded his first app to $25K/month in 5 months, and the app was the easy part. Umar couldn't code at all and built TradeZella into an 8-figure SaaS on the back of an audience he'd built for years. The builders aren't winning because they build. They're winning because people are already watching when they do.
Lazar's advice for non-technical founders
When Pat asked whether people should learn to code, Lazar's answer was blunt.
"By the time I become a mediocre code writer, AI will be so much better than what I am."
His advice instead: learn to vibe code. Learn how these tools think, how to structure context for them, how to talk to them. Learn how to prompt.
Six months ago I'd have pushed back on that. Now? He shipped 50 projects in 6 months with a day job. I know senior engineers who shipped less.
The stack he used, for reference:
- ChatGPT custom GPTs for the base prompt, the 4 planning docs, and debugging
- Lovable as the builder (chat mode 80% of the time)
- GitHub to own the code and store the docs
- Supabase for database and auth
- OpenAI API for the AI assistant feature
- 21st.dev for prebuilt UI components
FAQ
What is a vibe coding workflow?
A vibe coding workflow is a repeatable system for building software with AI tools instead of writing code by hand. Lazar's version: generate a base prompt in ChatGPT, turn it into 4 planning documents (master plan, implementation plan, design guidelines, app flow), commit them to GitHub, then build in Lovable while spending about 80% of your time in chat mode reviewing plans before any code gets written.
Can you really clone a $1M SaaS with AI tools?
You can clone the features, not the business. In the video, Lazar replicated Starter Story's case studies, ideas library, academy, community, and memberships pages, plus a working database and an AI chatbot. What he couldn't clone: 8 years of content, SEO rankings, an email list, and audience trust. Those are the actual moat.
What tools did Lazar use to clone Starter Story?
ChatGPT (with custom GPTs for planning and debugging), Lovable for building the app, GitHub for storing the code and docs, Supabase for the database and auth, the OpenAI API for the AI assistant feature, and 21st.dev for prebuilt UI components. Total cost is roughly the price of a few subscriptions, against a product that took thousands of hours to build by hand.
Do I need to learn to code before using AI coding tools?
Lazar says no, and he's shipped 50+ projects without writing a line. His argument: by the time you become a mediocre programmer, AI will already be a better one. The skill worth learning is structuring context: writing planning docs, giving screenshots as references, breaking work into small steps, and reviewing the AI's plan before it builds.
Steal the workflow, then build your moat
If features can be cloned in an hour, the founders who win are the ones with distribution and a real edge. That's exactly what I dig into on the Profitable Founder Podcast: bootstrapped founders making $100K to $10M a year, sharing the numbers and playbooks behind the part you can't clone.
