Vibe coding an AI ads creation tool is pretty cool! (Drop your experiences too!)
Hey everyone!
I've been building an AI marketing tool for almost a month now and wanted to share my experience so far.
I built Krezo, an AI marketing tool where you can paste your website URL into Zappy (my AI assistant), and it'll analyze the site and pull whatever useful information it can — logo, brand colours, what the business does, products, etc. Krezo then uses that information to create ads for the business. Right now it can generate image ads, headlines, descriptions, hashtags, different creative concepts, and more. I'm also working on video generation, but that's still a work in progress.
But the interesting part has actually been how I've been building it.
I've been using two different AI coding models: Codex and Claude Code.
And WOW. Using them together has made a much bigger difference than I expected.
I mainly use Codex for backend work like audits, generation pipelines, image/video generation logic, testing, API routes, product fidelity, etc. Then I use Claude Code for frontend work like UI/UX, mobile responsiveness, loading/error states, visual polish, templates, and making the overall product actually feel good to use.
At first, I was using Claude Code for pretty much everything. The UI was coming along great, but the actual generation side wasn't nearly as good. Some of the generated ads just didn't make sense, and I kept running into problems with how everything worked together behind the scenes.
Once I brought Codex into the project and started separating responsibilities between the two, it was honestly a game changer. Image generation has gotten WAY better, Zappy actually understands what it's supposed to do when editing creatives, product fidelity has improved, and the whole generation pipeline feels much more reliable.
The biggest thing I've learned is that if you're vibe coding a serious project, don't necessarily make one AI model responsible for everything. Giving different models different responsibilities has worked really well for me. They also catch things the other one misses. Obviously AI still screws things up and I have to keep an eye on what they're changing 😂 but it's kind of funny — sometimes the AI needs another AI to keep it in check.
Has anyone else been building this way? Curious if you've noticed certain coding models being noticeably better at frontend vs backend work.
And if anyone wants to checkout what I've built so far, Krezo is at trykrezo.com. It's still very much a work in progress, so genuine feedback would actually help.
Thanks guys!