How I work as an AI-assisted engineer

For ten months I built features on a production Ruby on Rails and React platform. My main engineering tool was Claude Code, with my own skills, hooks and sub-agents.

The tool wrote many first drafts. I owned every line that went to review. This post explains how I keep those two facts in balance.

The AI does not remove the review

A model writes code fast. A model can also write a wrong sentence in a confident voice. So I read every line of the diff before I push. I verify every message that a model writes before a person reads it.

AI-assisted engineering is not less engineering. It moves the work from typing to reading, proving and deciding.

Teach the rules once

I write skill files that hold my rules: code style, how to prove a change, and how to write for a customer. The model loads the skill, so it follows the rules without a reminder.

When the model makes a mistake, I save the correction as a memory. The same mistake does not come back in the next session.

Building features that call a model

When a feature calls a language model, I follow four rules:

• Send every call through one place, so you can change the model without a hunt.
• Track the cost of each feature on its own from the first day.
• Fail clearly when configuration is missing. Never fall back in silence.
• Check the output of the model before a person sees it.

What I learned

Review AI output before another person reads it. Give the model the rules, not reminders. Know what each AI feature costs.