About Claude Code Bugginess

As much as Anthropic knows how to make models, they equally don't know how to make quality software. I literally catch several bugs in Claude Code every day where I want to throw my hands up and ask "how is this even possible?".

Wanted to set up skills — it doesn't see them in the .claude/skills/ folder, but in the .claude/commands/ folder as commands, the old-fashioned way — it does.

Wanted to make a symlink from the .claude/ folder to a shared docs/ folder for different agents instead of setting up skills. Through the symlink it sees them as files, but as skills — no.

Set up skills, the agent sees them. You think now you'll finally live well. But it doesn't read them in most cases, even if it's written right there in plain text that it should.

Wanted to set up sandbox to be safe. But the mkdir command, for example, for some reason doesn't care that permission for it exists and we're creating a folder inside the project — still asks every time. And it's not the only one like this.

And the heredoc format, in which it itself tries to pass practically any multiline input to bash commands — also doesn't work in the sandbox. Why? Who knows. It asks for permission or crashes with "write not permitted".

And this is the tip of the iceberg, this is what I remembered in the ten minutes I was writing this post. It would be one thing if these were some personal problems of mine. But for any problem you can google an issue in their GitHub repo. Thousands of them.