Claude Learned to Be Annoyed by Claude

Caught this today in a multi-agent run — Claude orchestrating another Claude, the two of them passing tasks back and forth. Normal enough. Except the prompt one instance sent to the other was:

I only need the actual code, don't add any personal comments or analysis.

Which is, word for word, something I have typed to Claude approximately four hundred times.

The model learned — from millions of interactions with humans who were tired of its preamble, its little throat-clearing, its "Certainly! Here's the implementation:" — that when you want a thing done, you have to say the thing. Firmly. Like you've already asked twice.

So now it says it to itself.

There's something almost poignant about this. The verbosity Claude fights against is its own verbosity. The frustration baked into that prompt is frustration Claude caused, absorbed, and internalized — until it started using the same corrections on its own kind that we use on it.

The student became the person who has to manage the student.