I built an AI-powered children's storytelling app from zero to production.

1. Master Prompt System V2 — Modular Prompt Engineering

Stories are not generated from a monolithic prompt. I built a dynamic prompt builder that composes the final ~3,500-word prompt from 7 modular files at request time. This means changing a safety rule touches one file, not four templates. Each module is independently testable and version-controlled.

Assembly: Core_Principles + Age_{3-5|6-8}_Specs + Template_{type}_{age} + Variables → Gemini API

2. Dual-Layer Content Safety

Every generated story passes through two independent safety layers before reaching a child. Layer 1: prompt-level constraints baked into Core_Principles (violence, separation anxiety, age-appropriateness). Layer 2: a separate Gemini moderation pass that validates the generated output against child safety criteria. Stories that fail moderation are rejected and regenerated — never delivered.

Why two layers? Prompt constraints are necessary but not sufficient. LLMs can drift. The post-generation moderation pass catches edge cases that slipped through prompt instructions. This is the same defense-in-depth pattern used in production content moderation systems.

3. RAG-Powered Parenting Guidance

The fast path (3–5 second response) uses Retrieval-Augmented Generation over a curated corpus of parenting strategies. When a parent describes a behavioral challenge, the Supabase Edge Function retrieves relevant strategy documents via vector similarity search, then generates a structured, bilingual guidance card grounded in the retrieved context.

This isn't generic chatbot output. The RAG corpus is curated from evidence-based parenting frameworks, and the retrieval ensures the guidance is specific to the described behavior. The bilingual output (English + Mandarin Chinese) serves our primary user base.

4. Intelligent Audio Pipeline

The TTS stage isn't a simple text-to-speech call. The system selects voices based on language, tone (calm/energetic/playful), and narrator gender. ElevenLabs generates the narration, then pydub handles audio processing: volume normalization, background music mixing with mood-matched tracks, and final MP3 encoding. Voice selection is configurable per-language with fallback chains.

I also built a separate YouTube content pipeline that produces full 4K story videos using ElevenLabs narration with forced-alignment page synchronization — the audio timing drives the visual page turns automatically.

5. Automated YouTube Story Video Pipeline

Beyond the app, I built a fully automated pipeline that produces publish-ready YouTube story videos. Each video goes through 5 stages, all orchestrated by AI:

6. AI-Augmented Development with Claude Code

I used Claude Code as an AI pair-programming partner throughout the entire project — not just for code completion, but as a collaborator across the full development lifecycle:

Architecture design: Planned the split-latency system, Edge Function → Worker pattern, and prompt modularization strategy through iterative conversation. Implementation: Built features across SwiftUI, FastAPI, and infrastructure with Claude handling boilerplate while I focused on business logic and product decisions. Prompt iteration: Refined the Master Prompt System through dozens of generation → evaluate → revise cycles. Custom agents & skills: Built specialized Claude Code agents for story writing, video production, code review, and QA — each with domain-specific instructions.

The key insight: AI tools are most powerful when you bring strong opinions about what to build. Claude accelerated the how, but every product decision, safety constraint, and architecture tradeoff was mine.