Straktur

Phone rang. Customer asking where her order was.

Turned out Zapier dropped a handoff between two Airtable bases 3 weeks earlier.

No alert. No error. Just… gone.

The product had an 8-week lead time.

That was the moment I knew the stack had to go.

I run Mroomy - we design kids' rooms. 100+ suppliers, more than 2500 of projects, 50–100 products per project. Everything stitched together with Airtable + Interfaces + Zapier (and WooCommerce, Trello, Dropbox, GDrive, Typeform).

At one point we had:

• ~50k records

• 40+ Zaps

• multiple bases cross-referencing each other

It worked for 6 years.

Then it didn't.

We hit the limits all at once:

• Airtable wanted the next pricing tier

• Interfaces couldn't show what we actually needed

• Updating one product meant changing it in 3 places

• New Zaps had to defend against old broken Zaps

And the worst part - failures were silent.

The phone call wasn't even the worst incident that month. Just the one I remember.

So I started rebuilding. Custom app. Own database. Own UI. And since it's 2026 - I let Claude Code do most of it.

First 2 weeks felt like cheating.

Schema? Done. CRUD? Done. Basic flows? Done.

Then I did an audit.

In one app:

• 4 different form validation patterns

• 3 ways to fetch data

• 2 completely different error handling approaches

Same app. Same week.

Each file looked fine on its own. The whole thing was a mess.

That's when it clicked. It wasn't bad code. It was lack of constraints.

I'd been letting the agent decide architecture on the fly, and every session it picked slightly differently. Different patterns, different assumptions. I didn't notice until I zoomed out.

That's the part nobody talks about with AI coding. The first 70% is unreal - UI, forms, scaffolding, even non-trivial logic. Then you hit the last 30% and the agent doesn't know how you want things done, so it picks differently every time.

If you don't define patterns upfront, you're not moving faster. You're just deferring the mess.

What worked for me was writing the conventions down before generating anything. One way to fetch data, one form pattern, one error handling approach, one folder structure. Then feeding that to the agent every session.

Same prompts. Same features. Completely different outcome.

Curious if anyone else actually audited their AI-generated code after a few weeks. What did you find?