Skip to main content
Back to Blog
How Conrad Came Together: AI, Payments and Software Practice
AISaaSClaudeCloudflareStripeBuildingInPublic

How Conrad Came Together: AI, Payments and Software Practice

Ulrich Diedrichsen
Ulrich Diedrichsen
••12 min read

How Conrad came together with Claude Code, Cloudflare and Stripe, and what remained uncertain after the first prototype.

Update, November 14, 2026: This article describes a development stage from February. A payment function does not by itself prove customer purchases or commercial success. The project history shows design work and initial commits on February 5, with a Stripe live-billing commit on February 6. The earlier claim of going from idea to live in five hours is not sufficiently supported. Some historical cost and maturity figures cannot be independently verified here either.

The Starting Point

In early February 2026, I was pursuing an idea: An AI tool that analyzes social media posts before you hit publish. Not as a moralizing warning finger, but as a dramatic, British-polite analyst who paints vivid scenarios of what happens after posting.

"Conrad" — the Consequences Eradicator.

During development, Conrad gained a Stripe payment flow, cost protection and a Chrome extension. The project history records work across February 5 and 6.

I had never used Stripe before.


What the First Version Included

An early SaaS prototype with the following features described at the time:

  • AI-powered post analysis for 5 platforms (LinkedIn, X, Bluesky, TikTok, Instagram)
  • Post improvement mode — Conrad rewrites your post without losing your voice
  • Magic link authentication — no password, no OAuth setup
  • Credit system with 3 free analyses per day
  • Stripe payment function — 50 credits for €5 were planned; the original report described the payment flow as live. Purchases by other people are not established here
  • User dashboard with analysis history, credit balance, billing management
  • Community wishlist with AI moderation
  • Chrome Extension (Manifest V3) for 5 platforms
  • Bilingual from day 1 (German + English, native quality)
  • Cost protection — per-user daily limits, global budget breaker, request deduplication
  • Admin monitoring — real-time cost overview

In numbers: 70 source files, 8,300+ lines of code, 34 commits, 5 database migrations, 85 specification entities.


The Tech Stack

Cloudflare — The Entire Infrastructure From One Source

This was the most important architectural decision. No AWS. No Vercel. No separate database server. All Cloudflare:

  • Cloudflare Pages — Hosting and deployment. One wrangler pages deploy and the app is live. Worldwide. With automatic SSL.
  • Cloudflare Workers — The SvelteKit server routes ARE Workers. No separate backend server needed.
  • D1 (SQLite) — Relational database directly at the edge. Users, analyses, credits, subscriptions — all in D1.
  • KV — Key-value store for rate limiting, usage tracking, request deduplication, budget breaker.
  • Turnstile — Bot protection without CAPTCHA.

For this initial version I did not need to operate my own Docker or Kubernetes infrastructure. Configuration, monitoring, security and ongoing maintenance still require work. The original report estimated low infrastructure costs; a complete accounting across services is not available for this update.

SvelteKit — The Framework That Gets Out of the Way

SvelteKit with the Cloudflare adapter is the perfect combination. A monorepo where frontend and backend live side by side. Svelte 5 with Runes ($state, $derived, $effect) is reactive without boilerplate.

Claude API — The Brain

Conrad uses Claude Sonnet 4.5 as its analysis engine. I used direct fetch() calls for this prototype. That is not a general limitation: Cloudflare documents Node.js compatibility for Workers, subject to the APIs and package involved.

The trick is tool_use: Claude delivers structured scores (JSON) AND free text recommendations in a single API call. The user sees results within 2 seconds while Conrad is still writing.

The original report estimated model costs at around $0.02 per analysis. Without a complete accounting of payment fees, infrastructure and taxes, it does not establish a reliable margin.

Stripe — Payments Without Prior Knowledge

I had never used Stripe before. Claude Code walked me through it step by step.

Important: No Stripe SDK. Stripe's API is so clean that raw fetch() calls are completely sufficient.

Claude Code — The AI Development Partner

And here it gets meta: The tool that built Conrad itself uses Claude as its engine.

Claude Code wasn't just a code generator. It was an architecture advisor, a Stripe expert, a security auditor, and a pair programmer all in one.


What We Learned

1. Cloudflare Is a Solo Developer's Best Friend

D1, KV, Workers and Pages offered one way to start this prototype. Operating needs depend on the product and its requirements.

2. No SDK Is Sometimes the Best SDK

We used direct fetch() calls for this version. Whether an SDK is suitable depends on its dependencies and Workers compatibility.

3. Cost Protection BEFORE the First Paying User

Five protection layers:

  • Per-user daily limit (100/day)
  • Global budget breaker (500/day total)
  • Request deduplication
  • Fail-closed
  • Platform-specific character limits

4. tool_use Is the Game Changer

Claude's tool_use can return structured data. The application still needs to check that the output meets its requirements.

5. Personality Is the Product

The same score (shitstormRisk: 75) is boring as a naked value. As "Post this and your recruiter pretends not to know you by Friday" it's unforgettable.

6. AI-Assisted Development Is Not Autopilot

The crucial difference from "vibe coding": We created a complete specification with 85 entities BEFORE a single line of code was written.

7. Bilingual From Day 1

With i18n keys and Claude Code's native German language quality, it was almost free.


The Numbers

MetricValue
Development timePreviously stated as five hours; total not established
Source files70
Lines of code8,300+
Commits34
DB migrations5
Languages2 (DE + EN)
Platforms5
TypeScript errors0
CostsHistorical estimates; complete accounting not established here

The Stack at a Glance

Frontend:    SvelteKit 5 + Tailwind CSS v4
Backend:     SvelteKit Server Routes (= Cloudflare Workers)
Database:    Cloudflare D1 (SQLite at the edge)
Cache/KV:    Cloudflare KV
Hosting:     Cloudflare Pages
AI Engine:   Claude Sonnet 4.5 (via raw fetch, tool_use)
Payments:    Stripe (Checkout, Webhooks, Customer Portal)
Auth:        Magic Links via Resend (JWT, HttpOnly Cookies)
Bot Protection: Cloudflare Turnstile
Extension:   Chrome Manifest V3
Development: Claude Code (Anthropic CLI)

Conclusion

The combination of Cloudflare (infrastructure without overhead), Claude API (AI without complexity), Stripe (payments without prior knowledge), and Claude Code (development without waiting) is a paradigm shift.

This experiment produced a prototype with a payment function quickly. That alone does not show whether the product can be operated reliably or has found a market.

This doesn't mean AI development is trivial. You still need product understanding, security awareness, and the ability to ask the right questions. But the barrier between idea and product? It's almost gone.

The version described at the time was intended to analyze real posts and accept payments. Independent customer purchases and long-term operation are not established in this article.

A quick start, not yet evidence of a working market.

Tags:

AISaaSClaudeCloudflareStripeBuildingInPublic
Ulrich Diedrichsen

Ulrich Diedrichsen

Software architect & product builder

Professional software work since my computer science studies. Ex-IBM, ex-PwC. Now building products with AI in Hamburg.