Claude Code review: surprisingly good, still needs a human

Claude Code has surprised me with how good it is on coding projects. I would recommend it to every developer, while keeping a clear distinction between getting useful help and handing over judgement.

I have used it across multiple projects. Those projects are private, so I cannot name them or share their code here. This is a personal review based on that work, not a benchmark, and I am not going to invent a productivity percentage to make the recommendation sound more convincing.

Coding is where I like it most

Working with Claude Code on coding projects is the part I enjoy most. It has helped me on multiple projects, and my reaction has been consistently positive: it is surprisingly good.

That matters more to me than an impressive isolated example. A useful coding assistant has to fit into work that already exists, with decisions already made and a person responsible for the result. That is the setting in which I want help.

I cannot give readers a public repository or a detailed account of those private projects. So the honest limit of this review is that I can recommend it from my experience, but I cannot offer a reproducible comparison against another tool.

My recommendation is to try it on a real task you understand well enough to assess. Decide whether the result solves your problem. A response that reads well and a change you are happy to keep are different standards, and the second is the one that matters.

Cowork impressed me with office productivity

My positive experience with Claude extends beyond coding. It has also been useful as an office working partner: creating PDFs, proposals, slides, and similar material. I have worked on multiple proposals with it, but those are private too.

Claude Cowork has particularly impressed me with its ability to work with me on office productivity tasks. I want to name it separately because that experience should not be quietly attributed to Claude Code.

My recommendation to developers comes primarily from coding with Claude Code. My recommendation to office workers also draws on Cowork. Both have impressed me, but they are different parts of my experience with Claude.

For someone who spends much of their working day preparing documents, I think Cowork deserves a try. I would start with the work you actually do, rather than judging it entirely on someone else's demonstration. You already know what makes a proposal or presentation useful to its intended reader; that gives you a basis for assessing the result.

Where Claude has frustrated me

The most frustrating problem I have experienced happened in regular Claude chat, not Claude Code. While discussing my health, it repeatedly mixed up doctors and their specialties.

This was not a one-off in my experience. I already knew it could mix up people and context, so I was watching for it. Otherwise, I could have gone to the wrong doctor. I did not actually do that; I caught the confusion before acting on it.

I am keeping the people and medical details private. The issue here is the incorrect association between a person and their expertise. I am describing what happened in my conversations, not making a medical recommendation or claiming that every user encounters the same problem.

It would be unfair to call this a Claude Code coding failure. It happened in chat. But it does influence how much I rely on an assistant's account of a conversation.

A confident answer can still attach a fact to the wrong person. When the identity matters, I want to check the association explicitly. Reading an answer and finding it plausible is not enough.

My recommendation

I would recommend Claude Code to developers, especially those willing to work with it on a real project and inspect the result. It has been good enough across my private coding work to earn that recommendation.

I would also recommend exploring Cowork to people who do a lot of office productivity work. The document and proposal work has impressed me for different reasons from the coding experience, and it deserves to be evaluated on its own terms.

The context mix-ups have not erased the value for me. They have made the boundary clearer: I am happy to get substantial help from these tools, while keeping responsibility for the facts and decisions that leave the conversation.