One of the first things the instructors told us: Don’t fall in love with your app.

Agrim Singh, one of the workshop leads, said building is so fast now that it’s often smarter to start over than to patch a messy product.

He told us to “reframe” failure. Killing an app isn’t a loss — it’s feedback. The faster you discard weak ideas, the faster you land on one that works.

When I built a simple personal trainer app using Manus, it took about an hour. Most of that wasn’t me coding. It was me waiting for the AI to generate and deploy the web app while I refined prompts. The barrier to building felt absurdly low.

“What actually makes a vibe-coded app succeed?”

I put that question to Sherry Jiang, another instructor. With so many apps being spun up in hours, I wanted to know what separates something that sticks from something that gets deleted.

Jiang said it’s not about interface polish, but about identifying a real user and a need.

That’s why it’s exciting to see non-technical people build, she said. When the cost of building drops to near zero, the bottleneck shifts to ideas and lived experience. A feng shui practitioner — someone trained in the traditional Chinese practice of arranging spaces to influence luck and well-being — can build an app to make the practice more accessible to others. A food stall owner can create a tool to streamline operations. Those ideas don’t usually come from someone trained purely in software.

None of us in the room knew how to debug. For a non-technical person, staring at a wall of error messages is a nightmare.

The only way to learn debugging, the instructors said, is to do it.

Jiang studied business administration at UC Berkeley. It was through vibe coding — and constantly working with AI — that she reverse-engineered her way into understanding coding fundamentals.

“You learn by asking the AI to walk you through it,” she said.

When my personal trainer app threw an error while loading exercise cues, I copied the message, described what was happening, and asked the model to fix it.

It worked, but another error appeared. I clarified the issue and tried again.

When I probed the model about the error, I learned that the app was trying to pull too much information at once, causing the page to crash. That helped me to understand how the error happened in the first place.

One of the most important skills in building an app is getting the first prompt right.

Instead of asking the AI to code immediately, start by asking it to plan the app. Review the plan, then execute.

The instructors said first prompts usually fail because they contain incorrect information, are incomplete, or cluttered. “A bad line of plan has a massive cascading effect,” they added.

When users ask the AI to generate a plan instead of building immediately, the model is forced to think through architecture, features, and flow before touching code. That alone improves the output.

After reviewing the plan, the next smart move is to ask the AI to question your idea. This step pushes both you and the model to surface assumptions that weren’t spelled out in the original prompt.

You might suddenly realize you haven’t decided whether this is a web or mobile app. Whether it should pull from external databases. Whether user accounts are required.

Another great piece of advice: sequence your thinking and your prompts. Start broad, then narrow down.

The AI ecosystem is huge.

There are reasoning models like OpenAI’s ChatGPT, Anthropic’s Claude, and Google DeepMind’s Gemini. For images, there’s Google’s Nano Banana and platforms like Fal. For voice, tools like ElevenLabs.

Even within vibe coding itself, there are different builders, including Cursor, Lovable, and Emergent.

Don’t be loyal to a model. Be loyal to speed and outcomes, the instructors said.

New updates roll out constantly. Capabilities shift every few months. A tool might suddenly take over the tech world.

Knowing what each tool is good at and plugging them in accordingly could dramatically improve results.

A good vibe coder doesn’t just write prompts. They know how to orchestrate tools.