If you also write real product code with AI, this feeling will be familiar: a feature ships fast, but a few days later you change something else and the earlier thing quietly breaks.

Every release needs a human in the loop — deciding by hand which tests to run, which logs to read, which failures to ignore, and which ones must be fixed. There’s no real confidence underneath it.

I’m building a harness now, not out of engineering perfectionism, but because these pain points have shown up too many times. So I decided to do this in public — a Build in Public series with one theme: standing up a harness inside a real monorepo.

An edited transcript of my late-2025 team talk, “Vibecoding: A Breakout Guide for Programmers in the AI Folding.” It starts with Folding Beijing, walks through survey data from 37 colleagues on how AI coding tools are actually landing in the team, and looks at how to push individual efficiency from an average of 38% to 80%+ for everyone — covering prompt techniques, lean MCP setups, and a bunch of tricks you can only learn from practice.