Solo + AI Pair Programming Notes

What works

• Claude Code for architecture decisions and refactors
• AI-generated commits — explicit "what" + "why" beats my hand-typed messages
• Diff-driven code review — paste diff, ask "what would break?"
• TypeScript first → AI catches mismatched types in edits

What doesn't

• Letting AI design the database schema. Always wrong on edge cases.
• AI-written API endpoints without me reading them line-by-line. Auth bugs.
• Long-running agentic flows on uncommitted code. Burns context.

Cadence

• 90% of code lands via AI pair
• 10% I write because the AI keeps getting it wrong, and that's a signal the code is doing too much.