Pair Programming With Modern LLMs and Agents

I Stared at TypeScript Code for Hours

It compiled perfectly. Tests passed. The feature didn't work.

Then I saw it: one wrong character. A colon instead of an equals sign. Valid TypeScript. Plausible code. Completely broken. The AI had generated a type declaration instead of an assignment.

That's when it hit me: you're responsible for everything the AI writes. When demo day comes—when users discover bugs in production—they won't complain to OpenAI or Anthropic. They'll complain to you.

This book is about mastering that responsibility while shipping software 3-5x faster than you thought possible.

Real Numbers From the Trenches

In Q1 2026, I shipped:

•49 pull requests on GitHub
•95,000 lines of code
•Three separate features originally scheduled across eight weeks of sprints—delivered in six days

Not because I'm superhuman. Because I learned how to configure AI tools properly and let them handle the mechanical work while I focus on architecture and judgment.

Most of my team jumped at it enthusiastically. A few resisters are falling behind, visibly, asking us to slow down. That's the reality of the productivity gap opening up across the industry.

Boris Cherny, who architected Claude Code at Anthropic, ships 10-30 pull requests daily. He hasn't manually edited code since November 2025. That's not science fiction. That's what happens when you stop thinking of AI as a tool you use occasionally and start thinking of it as your default development environment.

What This Book Actually Teaches You

This isn't a collection of prompting tips or another "10 Best AI Tools" listicle. It's the complete workflow—from choosing tools that matter (most developers are still using glorified autocomplete) to configuring them properly so they understand your project, your conventions, and your constraints.

Configuration That Actually Works

•Rules files that teach AI your coding standards without you explaining them every time.
•Skills that automate entire workflows—ticket grooming, PR creation, code review—not just code completion.
•AGENTS.md files that give AI deep project context: your architecture, your conventions, your team's specific patterns.

Real Workflows in Action

•Shell scripting as conversation, not syntax memorization. Stop Googling flag orders. Just describe what you need.
•Git automation that handles merge conflicts, writes commit messages, generates PR descriptions from your actual changes.
•Test-Driven Development with AI that writes failing tests first, implements one at a time, commits each pass.
•Project management through AI that went from "thing I avoid" to "thing AI handles while I build."

Honest About What Goes Wrong

•The TypeScript bug that compiled but didn't work.
•The REST endpoint that included /api in the path—syntactically correct, completely wrong for our backend.
•The calendar component that never validated you can't schedule events in the past because nobody explicitly stated that constraint.
•Code that looks plausible but fails at runtime.