How I ship with AI agents

AI coding agents are fast and confidently wrong. The difference between “I used AI” and “I engineered with AI” is the setup around them. Mine has five parts, and I use it on every project.

1. Rules the agent reads first

Every project has one short instruction file: hard rules (what must never happen), a fixed vocabulary, where the source of truth lives, and the current state in about ten lines. When it goes stale it does damage, so keeping it current is part of the work.

2. Every decision written down

Every design decision is recorded with its date and reason. When a decision is reversed, that’s a new entry, not an edit. When any document disagrees with this record, the record wins. Agents and people both work from it, so nobody has to remember why something is the way it is.

3. Skills for anything I repeat

A release or a full verification run is one named command that encodes the exact steps, so I’m not re-explaining them every session. I keep only the skills I actually use. A hundred generic ones made the agent choose badly.

4. Hooks that keep artifacts honest

Small automations run when a session ends, for example regenerating a project tracker from its data, so status reports can’t drift from reality.

5. Verification before “done”

Nothing counts as done until tests, a browser walk-through, or a real command output says so. Deploys to shared environments need an explicit human yes, every time.

What this bought me

I built a multi-service production platform from design to a UAT deployment in about six weeks (case study), with every decision traceable.