Three weeks ago, I started using Claude Code properly as part of my design work.
Not to write production code. I still cannot tell you the difference between useEffect and useState without looking it up. I was using it to make things I could react to.
That is how I have always designed. I do not have the whole answer in my head before I begin. I need to push pixels, look at the screen, notice what feels wrong, and then move it somewhere better.
Claude gave me a different way into that loop.
I would ask it to build a first version. A page would appear in the browser. Then I would do what I always do. The spacing is too dense. The card is too heavy. The interaction does not feel right. I would say so. It would try again.
By the end of those three weeks, I had run 68 sessions and taken 88 screenshots. I had not expected the screenshots to matter that much. They were not proof that the agent was working. They were how I worked out whether it had made the right thing.
At some point, I realised I was doing the same thing after every task, even when the work was not visual.
The agent would tell me it was done. I would open the result anyway.
If it had changed a file, I wanted to see the diff. If it had sent something, I wanted to know what it sent and where. If I had asked it to tidy one folder, I wanted to know whether it stayed in that folder.
The word "done" was doing a lot of work there.
Sometimes it meant the task had run. Sometimes it meant the agent had stopped. Sometimes it meant it had done exactly what I meant. Those are not the same thing, and I kept catching myself treating them as if they were.
The agents I trust most are not the ones that sound the most certain. They are the ones that make their work easy to check.
That can be a diff. A short list of files touched. A link to the result. One line saying what I should look at before I move on.
“I changed X. Here is where to look. I am less certain about Y.”
That is more useful to me than a green checkmark trying to make me feel safe.
I want to be honest about what this is not. I do not want to review every tiny change. If I have to inspect everything, I have not delegated anything. I have just hired a very fast assistant who asks me to do the last mile every time.
But the bigger the consequence, the more I want the work to be legible. A spacing adjustment can happen quietly. A production change, an external message, or anything that touches more than I asked it to touch should leave a clear trail.
I used to think the hard part of working with an agent was getting it to understand what I wanted. That is still hard sometimes. But I am starting to think the more interesting part comes after it answers.
Can I see what happened? Can I tell whether it is right? Can I point it somewhere better if it is not?
That is where the work starts for me.