A small team's AI stack, honestly reviewed

Every few months someone asks what tools I use to build software mostly alone. The honest answer is boring in the best way: a short list, chosen by subtraction. Here it is, including the departures.

The principle: tools must disappear

A tool earns its place by vanishing from my attention. Every additional service is a login, a bill, an update, and a small tax on thinking. One person plus AI can genuinely do the work of a small team now — but only if the coordination overhead of the toolchain stays near zero. So the stack is graded on one metric: days per month I think about the tool at all.

What stayed in the stack

• TypeScript + Next.js. One language across the whole product. The framework's opinions replace a hundred small decisions, and static generation keeps the operational surface tiny.
• An AI pair, all day. Code, drafts, review, rubber-ducking. The workflow that survived a year of experimentation: I decide what and why, AI accelerates how, and everything it produces gets read before it ships. Unreviewed AI code is technical debt with excellent posture.
• PostgreSQL. Boring, forever, correct. Every time I've been tempted by something more fashionable, the temptation aged worse than the database.
• Tailwind + a token file. Design tokens in one place, utility classes everywhere else. It reads as controversial and works as infrastructure.
• Vercel. Deployment as a non-event. git push is the whole release process for most of my projects.
• A plain text file of prompts. The least glamorous item is the highest-leverage one: refined, reusable prompts for recurring work — code review, docs, test scaffolding. My prompt library outperforms most paid tools I've tried.

What left, and why

• The second AI coding tool. Two assistants meant two mental models and no compounding skill. Depth with one beats breadth with several.
• A separate design tool for solo projects. For products I build alone, designing in the browser with real data beats designing pictures of the product. (With clients, a design tool still earns its place — shared artifacts matter.)
• The project management platform. For a team of one-to-three, it was theatre. A markdown file per project, reviewed weekly, does the job with zero ceremony.
• Analytics beyond the basics. Page counts and Core Web Vitals answer every question I actually act on. The rest was dashboards as entertainment.

The economics of one person plus AI

The quiet revolution isn't that AI writes code. It's that the minimum viable team size for real software dropped to one. Design, build, document, support — each got compressed enough that a single disciplined person can hold an entire small product.

Three honest caveats from living this:

1. AI amplifies existing judgment; it doesn't supply it. My electronics-era debugging habits and teaching-era clarity do more work than any model upgrade.
2. The bottleneck moved to deciding. Building is fast now. Knowing what not to build is the scarce skill.
3. Solo doesn't mean isolated. Real users, early and often, replace the colleague who would have told me an idea was bad.

The stack fits on an index card, and that's the point. The tools disappeared. What's left is the work.