When I don't use Claude Code

For 23 parts I've been talking about everything Claude Code can do. Time for the honest counter-question I get asked in almost every conversation: "And where do you leave it alone?"

There are such cases. And knowing them matters at least as much as knowing the tricks.

🚫 Where I deliberately work myself

• A one-liner I already have in my head. By the time I've described it, I'd have typed it long ago.
• Security-critical core logic. Permissions, encryption, billing. Claude may suggest and review, but I consciously own the decision and every line.
• When I don't understand the problem yet. Then I very quickly get very plausible code for the wrong question.
• Production systems without a safety net. No direct access to live databases or servers. Claude works locally for us, in a branch, with review (part 7).

🧭 Where I'm careful

• New libraries and fresh versions. The model's knowledge has a cutoff date. Current docs belong in the context here, e.g. via MCP (part 5).
• Big rewrites in one go. Better in stages, with a plan (part 10) and tests (part 18).
• Learning. If you want to learn a language, write it yourself. Claude as an explainer, yes. As a ghostwriter, not so much.

A word of caution:

The line keeps moving. What I still did myself six months ago, I let Claude do today. That's why every few weeks I deliberately try a task I previously "didn't want to hand over".

My rule of thumb:

• Faster to type than to explain? Do it yourself.
• Don't understand it? Understand it first, ideally with Claude as a sparring partner.
• Would a mistake be expensive? Claude as a second pair of eyes, not as the author.