PromptGear
AI production platform for web & mobile apps
Hey IH 👋
Two months ago I mass-generated around 300 websites using my own AI builder. I'll be honest: the first 100 were so bad I wanted to mass-delete them and pretend this never happened.
But somewhere around website #100, something weird happened. People started saying "wait, AI made this?"
Here's the full story.
CHAPTER 1: The Confidence Phase
I launched PromptGear thinking I'd cracked it. Type a prompt, get a website. Great, right?
First user prompt: "luxury flower shop"
What I expected: elegant, modern, beautiful.
What came out: a website that looked like a funeral home designed by someone who just discovered the color purple. Every shade of purple. Purple gradients on purple backgrounds with purple text.
The AI had technically delivered. It was indeed luxurious. And it was indeed purple.
CHAPTER 2: The "Why Do I Even Bother" Phase
So I started testing obsessively. Every day, dozens of prompts. Here's a fun sample of what my AI produced:
"modern tech startup" → a page with so much whitespace it looked like a 404 error
"cozy coffee shop" → stock photo heaven, zero personality, Comic Sans somehow appeared despite not being in any prompt
"professional law firm" → actually decent, but it put "WE SUE EVERYONE" as the hero heading. Technically accurate for a law firm I guess.
At this point I had two options: quit, or figure out why every AI website builder produces the same generic garbage.
I chose to figure it out.
CHAPTER 3: The Dirty Secret of AI Web Builders
I tested every major competitor on the market. Most of them.
And here's what most of them do: they pick a template, swap colors and text with AI, and call it "AI-generated." It's basically a template engine with a chatbot on top.
That's not building a website. That's filling in a Mad Libs.
The real challenge isn't generating HTML. Any LLM can spit out a <div>. The challenge is making something that:
Actually looks like a human designer made it
Works on every screen size without breaking
Matches what the user had in their head (not what the AI hallucinated)
Doesn't look like every other AI website
And THAT took me 100 failed generations to understand.
CHAPTER 4: What Actually Fixed It
I'm not going to share exactly how (you're welcome, competitors 😘), but the breakthrough was realizing that AI doesn't need more freedom, it needs more constraints.
Think about it. When you tell a junior designer "make something cool," you get chaos. When you tell them "make a landing page with a dark background, one accent color, max 3 sections, no gradients, typography-focused," suddenly they produce gold.
Same with AI. The more specific the rules, the better the output.
After implementing this, generation #147 was a flower shop that actually looked like a real flower shop. Not a funeral home. Not a purple explosion. A clean, elegant, "I would actually buy flowers here" website.
That was the moment I knew this could work.
CHAPTER 5: The Numbers Nobody Talks About
Here's what indie hackers building AI products won't tell you:
Your AI costs are the tip of the iceberg. The API call is maybe 30% of your actual cost per user. The rest? Hosting, storage, deployment infrastructure, monitoring. I learned this the hard way after my first month's bill.
Free users are your best QA team. They will try to generate a clone of Apple, a PayPal phishing page, and a "full SaaS with Stripe integration" from a single prompt. Each of these broke something. Each made the product better.
Nobody. Cares. About. Your. Tech. I spent weeks building an elegant backend architecture. You know what the first user feedback was? "Can the fonts be bigger?" I could have been running this on a potato server and they wouldn't know the difference, as long as the output looked good.
CHAPTER 6: Where I Am Now
Today, PromptGear generates websites that actually look good. Not "good for AI," just good. A complete website, from prompt to a live page with its own URL, in 2-3 minutes.
And I'm not stopping. Every day I test the capabilities of different AI agents, compare their outputs, fine-tune prompts and push the boundaries of what they can do. One agent handles layout better, another is stronger at copywriting, another has a better sense of color. The art is in finding the right combination and making them work together so the result doesn't look like a compromise, but like intent.
It's not a "finished product." It's a constant experiment, and that's exactly what I love about it.
The Actual Lesson
Every AI product goes through the same cycle:
"AI is magic!" → you ship it
"AI is garbage" → users show you reality
"AI needs guardrails" → you figure out constraints
"AI + good constraints = actual product" → you finally have something
Most builders are stuck at step 1, selling "AI magic." The ones that win will be the ones that mastered step 4.
Try it yourself: promptgear.dev - you can generate a homepage for free, no signup needed.
Fixed a critical bug - guest usage limits were stored in Redis, which loses data on restart. Moved everything to Supabase so limits now persist across deploys. Also improved the visual editor - floating toolbars now look clean on both dark and light backgrounds. Organic traffic keeps coming in from IH, Product Hunt (#104), HN, and Facebook. Still zero ad spend.
Redesigned the homepage based on first-day feedback. Bounce rate dropped 35%. Added a "Built with PromptGear" showcase with live production apps. Also integrated Vercel Speed Insights for real user performance monitoring. All traffic so far is 100% organic - no paid ads.
Type a prompt, get a full React web or mobile app. Real code, ZIP export, deploy anywhere. White-label ready for freelancers & agencies. Built with Next.js, Claude API, Supabase. Live at promptgear.dev , no sign up, first 100 get a free generation.
-
1
cool build. one thing that usually decides retention for prompt-to-app tools is first deploy success rate, not codegen quality.
id track 3 numbers: time to first live deploy, percent of users who connect auth+payments in first session, and percent who ship a second app in 7 days. those 3 tell you if this is becoming workflow infra vs just a demo tool.
-
1
Really good point and exactly the kind of metric I need to focus on.
Right now, deploy to Cloudflare Pages is one-click, so time from prompt to live site is cca. 3 minutes. I'm tracking auth + Stripe connection rates but don't have enough data yet for solid percentages. The "second app in 7 days" metric is great :-) I wasn't tracking that specifically, adding it now.
This kind of feedback is exactly what separates a shiny demo from real workflow infra. Appreciate it.
-
-
1
Interesting direction. As a founder, tools that reduce build time are great, but the real value shows when they also help with iteration speed. Being able to test ideas quickly and refine based on user behavior has been far more important for us than just shipping faster.
-
1
Thanks Predrag! Totally agree, iteration speed is where the real value is. That's why we built a visual editor on top of the generated code. You generate the base, then tweak text, colors, images, layout all without tokens, unlimited free edits. So you can ship fast AND iterate fast based on feedback.
-
-
1
This is impressive — getting production-ready output instead of just demos is the real challenge.
Curious where you see most of the real use case today: freelancers spinning up client MVPs, or founders validating ideas quickly?
Also, how are you handling things like auth, payments, and environment config in the generated apps? That’s usually where “AI generated” projects break in real life.
-
1
Great question! Right now it's a mix of both. Freelancers using it for client MVPs and founders validating ideas quickly. The sweet spot is agencies who need to deliver multiple projects fast.
For auth and payments users can connect their own Supabase for user accounts and Stripe Checkout directly in the generated app. Environment config is handled through our dashboard, so the generated apps come ready to deploy with all integrations wired up.
-
-
1
Congrats! 🎊. Nice stack choice. Supabase plus Stripe plus Cloudflare keeps the surface predictable, and good to see you’re not storing card data directly.
Quick security questions:
Since you store generated websites and project content, how are you isolating projects between users at the database level?
Are Anthropic requests strictly server side with no API keys exposed to the client?
-
1
Thanks! Good questions:
Project isolation every project is scoped to the authenticated user's ID at the database level. Row Level Security (RLS) in Supabase ensures users can only access their own projects. Guest projects are stored client-side only (localStorage) and never touch the database.
API keys yes, all Anthropic requests are strictly server-side. No API keys are ever exposed to the client. The frontend talks to our Express backend, which handles all AI calls securely.
-
1
Great to hear you’re using RLS properly. Scoping everything to the authenticated user at the database layer is the right foundation. Keeping guest projects client side only also keeps the server surface smaller.
Server side only AI calls is the correct move as well. No exposed keys, no direct model access from the browser. That removes a whole class of abuse.
Since you’re already thinking this way, one thing we often see later is logic flaws between project IDs and user IDs in complex flows, not just direct reads. Worth periodically testing for IDOR and cross project access edge cases as features grow.
We’re a security team and built Nautillo Pro to simulate these kinds of real world attack paths against live apps. If you ever want an external check before a major release, you’re welcome to try our web attack simulator. There’s a free version for solo founders, so you can run a simulation monthly and stay ahead of surprises.
Nice work on the architecture. Keeping the surface predictable early pays off.
-
1
Thanks for validating the architecture - IDOR and cross-project access is exactly what I'm focusing on in the next security pass. RLS policies are scoped per user, but you're right that as features grow, periodic testing for logic flaws between project and user IDs is essential. Nautillo Pro looks interesting, I'll check out the free version. Appreciate the tip!
-
Comment