DEV Community

Cover image for WAVE Coding: Why we built 78 integrations for AI instead of letting AI build them
Michael Egberts
Michael Egberts

Posted on

WAVE Coding: Why we built 78 integrations for AI instead of letting AI build them

Every week I see another "I built a SaaS in 4 hours with AI" post. And every week, the comments are the same: "Cool, but does the Stripe integration actually work?"

Usually it doesn't.

That's vibe coding. You prompt, you hope, and you pray that the AI correctly implements a payment flow it's never actually tested. It hallucinates webhook handlers. It guesses at email configs. It builds checkout flows that break on the first real transaction.

We took the opposite approach.

The puzzle piece pattern
We're building WebsitePublisher.ai — a platform where AI assistants build and publish websites and web applications through conversation. Available on 9 AI platforms (Claude, ChatGPT, Gemini, Cursor, Windsurf, GitHub Copilot, Grok, Mistral, n8n) via MCP protocol.
In the last 16 days, we shipped 78 integrations. Each one is a self-contained puzzle piece — proven software running on our web servers. AI doesn't generate the integration code. AI calls the integration.
Here's what that looks like in practice:

User: "I need a webshop with Stripe payments and order confirmation emails"

AI selects:
→ product-catalog (MAPI entity + helpers)
→ shopping-cart (session-based)
→ checkout-flow (orchestration engine)
→ stripe (payment processing)
→ invoice-generator (PDF + accounting)
→ email-templates (Resend rendering)

Result: 6 tested puzzle pieces combined into a working application.

Zero hallucinated Stripe webhooks. Zero guessed SMTP configs. The heavy lifting happens in proven software on the server.

Why this matters
The fundamental problem with vibe coding is that AI is asked to do two things at once:

  • Understand what you want (AI is great at this)
  • Implement reliable infrastructure (AI is terrible at this)

WAVE coding separates the two. AI handles #1 — understanding your intent and selecting the right puzzle pieces. The proven software handles #2 — the actual Stripe calls, the email delivery, the database queries.

What's in the 78 puzzle pieces
Some highlights from what shipped in 16 days:
E-commerce stack:
Product catalog, shopping cart, checkout flow, order management, inventory tracking, shipping (MyParcel), invoice generation, discount codes
Communication:
SMTP email, email templates, contact forms, multi-layer spam protection
Data layer:
Server-side rendering for SEO, batch update/delete endpoints, data grids with validation
AI layer:
Coach (guided website creation), concept generation, streaming chat
Platform hardening:
DNSSEC, request tracing, error envelope standardization, security hardening

Each piece follows the same pattern: a handler receives the endpoint, input, and project ID. Dependencies are explicit. No magic.

The results
In the same 16 days:

  • 7 new customers onboarded
  • World Cup prediction game deployed for PSV Supporters (30,000 members)
  • Visual editor upgrade shipped
  • Coach AI guidance system improved

We're calling it WAVE coding

Not because it's a clever acronym. Because each application you build is a wave — one deliberate push that combines existing puzzle pieces into something new. Each wave builds on what came before.
Vibe coding is random energy hoping to land somewhere useful.
WAVE coding is deliberate momentum. 🌊

Curious what you think. Are you building infrastructure for AI to use, or letting AI build infrastructure from scratch?
websitepublisher.ai

Top comments (0)