Working with an AI coding agent: what actually works

The website you are reading was built with an AI coding agent, from design and code to tests and going live. The agent wrote most of the code. My part of the work did not disappear; it moved somewhere else. Here is what I found actually works.

Write down what you want before any code

An agent does its best work from a clear description: the goal, the limits, what “done” means. A few short paragraphs on who the page is for, what it needs and what it does not need save hours of back and forth. If you cannot describe it yourself, the agent will not guess it either.

Small tasks, each one checkable

“Redo the whole interface” is a bad request. “Make the GitHub grid blue, fill the card, and add a test that checks the colours” is a good one. Small tasks are easy to review, easy to undo, and bugs show up early.

Tests are the contract

The way I trust code an agent wrote is tests: write the test first, watch it fail, then make it pass. Tests turn “looks right” into “proven right”. When the design or the content changes, the old tests point straight at what was affected.

Look with your own eyes, not just the tests

Green tests do not mean it looks good. More than once every test passed, yet the text was too big, the spacing too tight, or an effect covered the words. Taking screenshots, checking on a phone, comparing with a site you like: that part is still the human’s job.

Keep the decisions on your side

The agent proposes things very quickly; the decisions stay with you: the design direction, which data goes public, when to go live, where secret keys live. A good agent asks before anything hard to undo. A good user does not skip those questions.

In the end

Working with an AI agent is not handing off a task and walking away. It is like working with a very fast, very diligent teammate who needs you to set the direction and check the work. When that works, one person can build what used to take a whole team.