I spent a second night in a row with Claude Code. There were still mistakes, but I was enjoying it enough to keep going. My immediate impression was that many problems came back to how I had described the task. I was also checking everything carefully.
That combination is more useful than either extreme of trusting every change or dismissing the tool after the first error. A coding agent can do a lot of work quickly. I still need a way to decide whether it did the work I actually wanted.
A request such as "fix the refresh" contains several possible tasks. Does the data fail to update? Does the display keep an old value? Does returning to the screen reset something unnecessarily? A person who has been staring at the application may think the answer is obvious. The repository does not contain the contents of that person's head.
A clearer request describes the observable behavior. It gives the starting state, the action and the expected result. That does not require writing a huge specification. Sometimes one concrete example removes more ambiguity than a paragraph of architectural vocabulary.
Review has a similar shape. I want to understand which files changed, why they changed and what evidence shows the problem is resolved. A passing build is useful, but it does not prove that the right behavior changed. A test that simply repeats the implementation can also be reassuring without being informative.
The Claude Code workflow makes it possible to move between reading, editing and running commands within the same conversation. That continuity is the attraction. It also means I should pay attention when a small request starts spreading into unrelated parts of a project.
I ran into a couple of timeouts during these first sessions. They were an interruption, not a reason to infer anything about the correctness of the code. Keeping the work understandable and recoverable matters even when the model is doing well.
What I liked after two nights was the pace of reaching concrete results. What I wanted to preserve was my ability to explain those results. A good session should leave me with a better application and a clearer understanding of the change, rather than a large diff that I am afraid to touch.