Building AI Tools with Lovable, Cursor, and Vibe Coding
I sat down with David Berkowitz on AI Insiders to talk about how I build AI tools as a non-coder. The stack I use. When to reach for Lovable vs Cursor. The meta-LLM trick that unblocks me. And the part nobody talks about: when you have to jailbreak your product off the platform that built it.
AI Insiders, by AI Marketers Guild
AI Insiders is the podcast and live series from the AI Marketers Guild. The show is hosted by David Berkowitz. The conversations are about how marketers actually use AI day to day. Not the hype. The receipts. Below are the moments from this episode I keep coming back to.
Looking for the gold or selling the shovel
“Either I am looking for the gold or selling the shovel.”
When David asked me to introduce myself, I gave him the frame I keep coming back to. My big agency years at WPP, JWT, and 360i / Dentsu were me looking for the gold: out there hunting, building campaigns that won awards and moved sales. My MarTech years that followed were me selling the shovel: making tools so other marketers could find their own gold.
Vibe coding has collapsed the distance between those two. I do both in the same week now. Day job at Hormel running social and influencer for Applegate, Spam, Justin’s Peanut Butter and the rest, then nights and weekends on ventures like JingleMyBrand and Social Lollipop where I’m back at the shovel. The two feed each other. The brand work tells me which tools the industry actually needs. The tool work makes me a sharper operator on the brand side.
I am not a coder. I am a producer who finally caught up
“I’ve always built things, but I’m not a coder. I’ve always had a lot of ideas. Pre-vibe coding, those ideas died because the resources cost too much and took too long.”
Years before vibe coding had a name, I built one of the first ChatGPT-powered WordPress plugins. It was called AI Prompt to Product. You could write a prompt, point it at a form, and turn the combination into a live feature on any WordPress site. Recipe in, gluten-free version out. Resume in, marketing-flavored version out. It worked. People paid for it. It also broke every time the OpenAI API changed, and every fix meant paying a developer. I was making money and losing money on the same product.
What changed in the last year is that the iteration loop closed. With Lovable, Cursor, Base44, Replit, and Claude Code, I can ship a tool in a weekend and fix it in twenty minutes when something breaks. The bottleneck used to be implementation. Now it’s judgment and the willingness to actually launch. I’ve shipped more tools that matter in the last six months than in the previous six years combined.
The actual stack: Lovable, Cursor, Claude Code
Hides backend complexity. Best for nights and weekends when you just want to ship. Costs more, but the cost of not shipping is much higher. JingleMyBrand was built almost entirely here.
Faster, cheaper at scale, more powerful. Expects you to be comfortable in a terminal and to wire up APIs yourself. My preferred environment when I have the time.
Base44 is more sophisticated than Lovable and a bit pricier. Replit is well-suited if you want a more developer-flavored experience without leaving the browser.
People get hung up comparing prices. The honest answer: all of these are rounding errors against what custom development used to cost. JingleMyBrand would have been thousands of dollars and several months of dev time on the old playbook. On Lovable, at the higher tier, it was a single weekend of focused work. If you spend eighty percent more on the platform that gets you to launch, that’s a deal every time.
My advice: pick one and ship a real thing with it before you start comparing. You’ll know which platform fits your brain after the first product, not before.
The meta-LLM trick that unblocks vibe coding
“Have a running conversation with ChatGPT or Claude on the side. Okay, now I’m trying to do this in Lovable. How do I do it. That meta layer.”
When David and I both agreed on the same trick, I knew it was worth saying out loud. Vibe coding platforms are powerful, but they aren’t magic. The moment you want to do something the platform doesn’t want to do (connect a niche API, handle an auth edge case, debug a build error), you stall.
The fix is to run a second AI in parallel. Open Claude or ChatGPT or Gemini in a separate tab. Describe what you’re trying to do in your platform. Ask for the specific approach, paste the error, paste the screen. The meta layer is your debugging partner. The vibe coding platform handles the build; the conversational LLM handles the strategy.
That dual setup is also how I avoid getting trapped inside a single platform’s worldview. The meta-LLM doesn’t care which tool I use. It cares about the problem.
Eventually you have to own the tech
“If JingleMyBrand gets big enough, I’ll jailbreak it. I don’t love that another company is essentially owning my technology, my platform, my IP in a sense.”
This is the caveat I always add when I recommend Lovable. It’s the right starting line, not the finish line. Your code, your hosting, and effectively your IP all live behind another company’s gate. That’s fine when the goal is shipping fast and finding out whether the product matters. It stops being fine the moment the product is earning real money or carrying real customer data.
The exit plan is simple. Export the codebase. Host it yourself on Netlify or Vercel. Wire up your own auth and billing. Walk away from the vibe coding platform once you don’t need the training wheels anymore. I haven’t done that yet for JingleMyBrand or Social Lollipop. I will when the numbers say it’s worth the effort. That’s the right way to think about platform debt: as deferred work, not free work.
Building something where AI in marketing comes up? I help teams cut through the noise and figure out what’s worth their time. Reach out and tell me what you’re working on.
Related reading
FAQ
Vibe coding is building working software by describing what you want in natural language and letting AI assistants like Lovable, Cursor, Claude Code, or Base44 generate, run, and fix the code. You operate at the level of intent and feedback instead of syntax. It doesn't mean unsafe or sloppy. It means the human stays in the design loop, and the model handles the implementation grind.
Lovable hides more of the backend complexity, so it's the right starting point if you have ideas but limited coding comfort. Cursor (paired with Claude Code) is more powerful and cheaper at scale, but it expects you to be comfortable in a terminal and to wire up APIs yourself. The honest answer: pick one, learn it well, and ship a real thing before you compare. The cost difference is rounding error compared to the cost of not building.
Start with a real problem that bothers you weekly. Pick one vibe coding platform (Lovable for ease, Cursor for control). Describe what you want; let the model build it. When you get stuck, run a separate Claude or ChatGPT conversation in a second tab as a meta layer, and ask it how to do the thing inside your platform. Connect APIs as needed for whatever the platform can't do natively. Ship. Iterate. The first version takes a weekend if you stay focused.
Initially yes, and that's fine. Lovable and similar platforms are how you ship in days instead of months. The tradeoff is that your codebase, hosting, and IP all live behind their gate. Once a product earns its keep, the move is to jailbreak it: export the code, host it independently (Netlify, Vercel, your own infra), and own the stack. Until then, the cost of platform lock-in is much smaller than the cost of not shipping.
Marketing Strategist & AI Expert