I'm building Tilzix, an app that turns your website or Instagram into carousels in your own brand. I use AI to write most of the code. I decide what gets built, test it, and own every bug that ships.

It passed my tests. It worked on my phone. Then strangers started using it, and every week they found something the AI and I had both missed.

Here are seven from the last two weeks, with the numbers. None of them were hard to fix. All of them were invisible until a real person hit them.

1. It guessed the brand's Instagram handle from the web address

To put the right @handle on each slide, the code took the first part of the domain:

handle: "@" + host.split(".")[0] 

Fine for glossier.com. Then a real customer pasted pos.qwin.co.ke, and their carousel went out with @pos on every slide. Another pasted a site hosted on Railway and got @railway.

I spotted it the same day in my analytics, emailed him, and fixed it that night.

The fix: read the Instagram link the site actually has, and if there isn't one, take the brand's name, not the host's.

const instagram = await siteInstagram(url); // the link on their own page if (instagram) brand.handle = `@${instagram}`; else if (isHosted(host)) brand.handle = `@${slug(brand.name)}`; // web-clipscript.up.railway.app → @clipscript else brand.handle = `@${mainName(host)}`; // pos.qwin.co.ke → @qwin 

That one line looked fine to me too. Neither of us asked: what does a real domain look like?

2. "Works on five networks", followed by four

A visitor made a carousel for a video tool. Slide two said "Works on five networks" and listed YouTube, TikTok, Instagram, Facebook.

Threads was missing. My slides allow four points, so the fifth was cut, and the headline still promised five.

Now every list headline is checked. If it promises a number that doesn't match the points shown, the writer has to fix it, by combining points ("Facebook and Threads") or by dropping the number.

3. The "are you human" check inside Instagram's browser

I use Cloudflare's bot check before anything costs money. In a normal browser it's invisible. Inside the Instagram and Facebook apps' built-in browsers it sometimes never loads.

So my page told people to "tick the box". There was no box.

One user in Lagos hit it four times in a row. The code gave up on the first error, even though Cloudflare retries by itself and often passes a second later. Now it waits for three errors, and in-app browsers get a hint to open the page in Safari or Chrome.

4. 887 animations running at once on phones

My home page shows live sample slides. Each one animates. I checked: 887 animations were running at the same time.

That's a lot of work to ask of a phone, for slides nobody could even see.

Now samples only run while they're on screen. It's about 9 animations most of the time.

5. My analytics counted Meta's robots as visitors

I ran a tiny Instagram ad. Day one: 15 visitors. I was excited.

Then I looked at them one by one. 8 of the 15 arrived within the same minute, from Prineville, Forest City, Fort Worth and Luleå. Those are Meta's data centres, checking my link when the ad was approved.

No dashboard told me that. I only saw it because I opened each visitor.

6. One event that ate my analytics plan

A "Love how it looks? Download it" prompt fired an analytics event when a carousel finished playing. The carousel loops. So it fired on every loop.

777 of 1,301 custom events in my account came from that one line, and the plan ran out.

7. The fix that never shipped

After fixing the @pos bug, a new user still got the wrong handle the next morning.

The code was right. But reading websites runs on a separate AWS Lambda, and that hadn't been rebuilt in three weeks. My deploy only updated the main site. The fix was sitting in a file nobody was running.

What I took from it

AI writes the happy path. It's very good at code that works for the example in front of it. It doesn't ask what a Kenyan domain looks like, or what happens inside Instagram's browser, or whether a loop fires an event a thousand times.

Tests didn't catch any of these. Every one passed. The bugs were in assumptions, not logic.

Watching real users did. I read my analytics one visitor at a time. Not the totals. The timeline: they pasted this, waited this long, tapped that, left here. Almost every bug above came from one person's trail.

So now my loop is: ship, watch five real people use it, fix what they found, repeat. AI makes the fixing fast. It doesn't make the finding happen.

If you want to see what the app does now, it's at tilzix.com: paste a website and you get a carousel in about a minute. Each of these was sorted within a day of someone finding it.

What's a bug AI wrote for you that only a real user found? I'd love to hear yours.