After using Claude Code every day for 10 months, it still surprises me.

In my last Advanced Claude Code workshop, Claude Code outright refused to follow the instructions in CLAUDE.md. It happened to 2 out of 16 learners.

Here is the setup. CLAUDE.md can live in different subfolders, and each one applies to its own folder. To demo this, we created two subfolders: english-1 and english-2.

In english-1, the CLAUDE.md said “Always write in Singlish.”

In english-2, “Always write in UK English.”

I named the first folder english-1 on purpose, not singlish. I wanted to prove the output is shaped by CLAUDE.md, not by the folder name.

We then prompted Claude Code to:

“Write an article about AI use cases in [industry] and save in @english-1/”

It worked for most learners. But for 2 of the 16 learners, Claude Code did the opposite. It explicitly said that while CLAUDE.md asked for Singlish, the folder is named english-1, so it decided to write in proper English.

One rule I always share about CLAUDE.md: always loaded does not mean always followed, especially once the context window fills up. But this is the first time I have seen Claude Code outright refuse and explain why.

The lesson: an LLM weighs everything in its context, not just the instructions you give it. When your context is confusing, the folder name conflicting with CLAUDE.md here, you are leaving it to luck which signal Claude Code follows.

If you are building with Claude Code, writing good instructions is only half the job. The other half is removing the signals that contradict them.

#AI #ClaudeCode #VibeCoding #AIAgents #BuildInPublic

If you are a non-techie who wants to learn Claude Code from a practitioner, I run a Foundations of Claude Code workshop.

Details: https://lnkd.in/e9r8kScw

If you are already familiar with Claude Code in the Terminal, my next Advanced Claude Code workshop is on 24 Jul, 10am to 6pm. This CLAUDE.md demo is one of the things we do there. We deep dive into context engineering, skills and workflow patterns to run big and long-running tasks (a.k.a. Loop Engineering).

Details: https://lnkd.in/eR6WPQnC