Earlier this month I was playing with one of Claude's "inspiration" Artifacts, a bedtime story generator (a React component that runs in the browser). I showed it to a mom I know. She said "This is nice, but it would be great if it could make my child the hero of the story and narrate it so we can listen together."

That offhand comment launched a 2-weeks learning journey from a simple React component to a full production SaaS.

This technical journey taught me that "just an LLM wrapper" is like saying a restaurant is "just a kitchen." Here's how I built it:

Stack: FastAPI + SQLAlchemy + Auth0 + Stripe + OpenAI API

  - Multi-step wizard flow with session management
  - Background audio generation with OpenAI TTS
  - User auth, subscription tiers, usage tracking
  - Rate limiting, input sanitization, secure file storage
  - Database migrations, monitoring, deployment scripts

The non-obvious challenges:

  - Prompt engineering to make the child the hero of the stories
  - Voice selection and prompts to find a tone that doesn't terrify kids at bedtime
  - Session state across wizard steps without losing progress
  - Handling OpenAI API failures gracefully
  - Storage for generated audio files
  - Making it parents and grandparent-friendly

Claude Code helped throughout, as a senior engineer to bounce ideas off and to write quite a bit of code. I reviewed, understood, and sometime rewritten every block of code. The learning was the point, in this case.

The humbling realization: Yes, it's "just" an LLM wrapper. But so is ChatGPT. The value isn't in the AI call - it's in solving a specific problem for specific people and package it in a user-friendly way. Parents don't care about tech stacks. They care that bedtime takes 20 minutes instead of 2 hours.

Site: https://bedtimestories.pro/ (3 free stories, then paid)

Feedback is welcome.

Anthropic settles class action from US authors alleging copyright infringement