When Claude Doesn’t Get It, the Problem Is You

I spent over two hours fighting Claude Code over a broken MCP config. Turns out I was the one who broke it.

Today I was using Claude Code CLI. I asked it to add an MCP server, and it did. Then I wanted to add another one, and it did. The issue arose when I asked it to create one at the global level, but the one it created initially was at the project level. They were out of sync.

It had two different configuration files for setting up MCP servers. After having Claude do some troubleshooting and resetting of configuration files, I finally got frustrated and told it to “Review everything and every setting,” and it finally found the extra configuration file and resolved my issue.

The Real Problem

Claude CLI was acting like a junior-level employee and only knew what it knew and only checked where I said to check. I was handing off the thinking and problem solving to AI.

The fault is all mine. I told it to update those files. But I did not understand the system well enough to tell it exactly where to look and what to change.

AI Is an Employee, Not a Partner

I manage the system that Claude is helping me run. I should treat AI like an employee. I have some onboarding… but sometimes I just let it go and do its thing. Then, when it does something I do not like, I tell it to fix it. Sound familiar? If you’ve used Claude Code for more than a week, you’ve done this.

What I Changed

I started being more explicit with my instructions. Instead of saying “add an MCP server,” I tell Claude exactly which config file to update and where. I think through the process before I let Claude handle it. If I do not understand how something works, I have no business delegating it.

After a change, either I or Claude verifies it works. Then I iterate on my instructions based on what happened. The command that finally fixed my MCP issue was telling Claude to “review every setting.” That should have been my first instruction, not my last resort.

The Takeaway

My key takeaway is to thoroughly plan the process before delegating it. Understand your system. Know where things live. Then tell your AI exactly what to do, like you would a new hire on their first week.

Next time Claude breaks something, or your junior-level employee does, ask yourself: did I actually tell it what to do? Or did I just hope it would figure it out?