An AI Coding Workflow That Keeps the Human in Charge

AI coding assistants are genuinely useful and genuinely dangerous. The difference is the workflow around them. Here is the one I use — it gets the speed and keeps the human in charge.

The rules

1. Write the plan first. A short PLAN.md — what moves where, which tests must pass, what success looks like. This is the source of truth you keep re-pointing the model at. Without it, long sessions drift.
2. Review every diff before accepting. The model is fast and confident; it is not self-reviewing. Read what it proposes, and reject anything that changes observable behavior without explanation.
3. Keep the test suite green after every step. If a change breaks a test, that is the signal — not the model's "looks good."
4. Treat it as a fast pair-programmer with a bad memory. It forgets earlier decisions in long sessions. Re-state constraints; do not assume it remembers.
5. Use it for the mechanical 80%. Refactors, test scaffolding, boilerplate — excellent. Architecture and "is this right for our users?" — that is yours.

Why the plan matters most

The single biggest failure mode is context drift: twenty minutes in, the model is confidently building something subtly different from what you asked. A written plan that you keep re-anchoring it to is the cheapest fix. In my experience it turns "impressive but wrong" into "consistently on-target."

The verdict

An AI coding assistant can compress a two-day refactor into a day — if the human does the reviewing and the tests stay green. The tool is not a replacement for understanding the codebase; it is a way to move faster through the parts you already understand.

Is it paying for itself? The calculator below lets you plug in your rate and the hours it saves — decide with numbers, not vibes.