From 5 Standalone Apps to a Unified Platform in 29 Days
In Part 1, I explained why I decided to build a fortune-telling platform with zero domain knowledge. Now let me tell you how it actually happened -- the technical build, the architectural decisions, the velocity data, and the AI-augmented workflow.
Phase 1: Monorepo Consolidation (Days 1-4) -- 188 Commits
Five standalone Next.js apps, each in its own repo. Day 1 was Turborepo setup and migration. 21 commits just to get everything building. Import paths broke. Tailwind configs clashed.
Day 2 was 38 commits of CSS restoration hell. Day 4 was the big win: the unified API layer with 52 routes migrated into a consistent pattern. Claude Code migrated all 52 in about an hour -- would have been a full day manually.
Also shipped: GDPR compliance, bilingual legal docs, and a DFA-based sensitive word filter.
Phase 2: Testing, Branding, and Identity (Days 5-10) -- 247 Commits
Vitest across the monorepo. GitHub Actions CI pipeline. Then the fun part: the PanPanMao cat brand emerged. Credit currency became dried fish treats. Landing page with "Thousand-year wisdom x Modern AI" tagline. Stripe integration went live.
Phase 3: Monetization and New Features (Days 12-19) -- ~300 Commits
The credit economy got serious: anonymous-to-authenticated merging, cross-tab credit sync via Supabase real-time, 7 contextual upgrade triggers, referral system.
The technical highlight: palm and face reading with MediaPipe running in-browser (21 hand landmarks, 468 face landmarks), canvas overlays at 30fps, multimodal AI prompts. Four days from zero to production.
Landing page redesigned from warm brown/gold to dark luxury. Users described it as "trustworthy."
Phase 4: The Milestone Rush (Days 20-27) -- ~400 Commits
Three milestones in four days. M1: Credit economy foundation. M2: Daily Hub with pre-generated content via midnight cron. M3: PostHog analytics.
Peak day: 98 commits. Three new products (child naming, compatibility analysis, annual forecast) from zero to production. By the third product, the pattern was mechanical -- the compounding effect of good architecture.
The AI-Augmented Workflow
My role shifted from "writer of code" to "director of code." I architected, reviewed, tested, and course-corrected. The AI was the execution engine.
Conservative estimate: 97% of the code was AI-assisted. But "AI-assisted" does not mean "AI-autonomous." Every line passed through my review. Every architectural decision was mine.
Surprising insight: AI was more consistent than I am. Claude applied patterns more uniformly to the 50th endpoint than I would have.
The Velocity Numbers
Metric | Value |
Total commits | 1,134 |
Average commits/day | 39 |
Peak day commits | 98 |
Product verticals | 9 |
API endpoints | 85 |
Approximate LOC | 284,000 |
Zero-commit days | 0 |
The shared packages were the secret weapon. When adding a new vertical, auth, credits, AI model selection, content filtering, streaming, error handling, and UI components come for free. The marginal cost of a new product is almost entirely prompt engineering.
In Part 3, I share the honest lessons -- what surprised me about product work, what I got wrong, and what this taught me about the future of solo building.