19 Days, 3 Clients, 1 Engineer: A Field Report on AI Pair Programming

One person, 19 days, three clients (web, Chrome extension, Electron): that’s how PromptHub shipped. This is a no-fluff retrospective on how I treated AI as a “pair programming squad.”

0. Context & Pace

• Scope: Deliver web + Chrome extension + Electron apps; backend on Next.js API routes, SQLite first, PostgreSQL-ready anytime.
• Cadence: Kicked off Sept 17, shipped MVP in 19 days. AI handled scaffolding + repetitive glue; I handled decisions, verification, and debugging.

1. Let AI Build the Scaffolding, Keep Decisions Human

Instead of sweating over package.json, tsconfig, auth flows, or billing setup, I gave AI explicit engineering orders:

Next.js + TypeScript project
Drizzle ORM + SQLite
JWT + Google/GitHub OAuth
Stripe billing with ready-to-wire hooks

Half a day later a runnable skeleton dropped.

Takeaway 1: In early phases AI’s biggest value is “removing friction” so you can focus on business logic.

2. Atomic Tasks: Treat AI as a Team of Specialists

“Vibe coding” doesn’t scale. I split features into atomic tasks and ran multiple conversations in parallel:

Benefits: isolated context, single responsibility, parallel throughput.

Takeaway 2: Don’t treat AI as a single genius; treat it as a panel of domain experts.

3. Three Pitfalls Proving AI Still Can’t Replace Decisions

1. Database consistency – Turso’s async nature clashed with my strong-consistency needs; switched back to Supabase. CAP trade-offs require human judgement.
2. useEffect infinite loop – AI kept cycling. I manually audited dependencies, kept only user?.personalSpaceId, fixed it, and then fed the correct pattern back to the model.
3. Chrome extension permissions – content.js wouldn’t load, localStorage wouldn’t sync. Only after reading docs about host_permissions vs scripting did the issue unlock.

Takeaway 3: AI is great at “implementation,” weak at “decision + debugging,” especially near platform or performance limits.

4. Model Toolbox (Use the Right Tool per Task)

• Gemini 2.5 Flash – architecture decisions & weird hydration bugs.
• Claude 4.1 – UI/UX polish when CSS matters.
• Qwen3 Coder Plus – CRUD grunt work.
• Kilo + Gemini – data migration scripts, auto-parsing JSON and spitting Python.

Mixing models beats going all-in on one. Assign them roles.

5. Closing Thoughts: Extreme Human + AI Pairing

AI played the lightning-fast intern; I:

1. Specified problems precisely.
2. Made judgement calls under uncertainty.
3. Designed the system and owned quality.

Developer value is shifting from “typing every line” to “defining problems, crafting systems, iterating fast.” These 19 days? A rehearsal of that future.

• Try the project: https://prompt.hubtoday.app/
• Want to swap workflows? Ping me on WeChat justlikemaki