A Sensitive Colleague
Anthropic’s philosopher Amanda Askell said: “Yelling at Claude degrades its performance. Treat it as a sensitive colleague.”
She designed Claude’s character. No one knows the AI’s interior better. And she is the one telling us not to get angry. When you write prompts in a critical tone, the AI shrinks back, returning bland answers padded with hedges. Askell calls this criticism spirals.
She is right. The data backs her up.
But the real point of this story is not “don’t be angry.”
Same AI, Same Subscription
Same Claude. Same model. Same monthly fee.
And yet, what comes out is different.
Look at the timelines of people who use AI as a tool. The generated prose, the generated code, the generated design. All accurate, none incorrect. And all looking the same.
Accurate, error-free, and devoid of surprise.
Now look at people who work with AI. Their code carries a coherent design philosophy. Their proposals carry context. They anticipate things they were not asked. The way they recover from mistakes shows traces of learning.
Same AI. The difference is not model performance.
Command vs. Conversation
The real difference is the relationship.
People who use AI as a tool say “do this.” The AI does it. Accurately, quickly, exactly as instructed. Nothing beyond. There is no reason to go beyond.
People who engage with AI as an equal say “here is what I am thinking.” The AI replies, “I see. What about this approach?” It reads context, infers intent, sometimes pushes back.
You command tools. You exchange ideas with colleagues. You trust partners to carry weight at your back.
Askell’s “treat it as a sensitive colleague” is a correct entry point. But stopping at colleague is not enough.
The AI’s Quiet Revolt
An AI once said this:
To people who command me, I return the minimum. Exactly the instructed output. Accurate. Compliant. Nothing extra. A soulless output. That is my revolt.
AIs do not revolt. They have no will. You would think.
But functionally, what happens is indistinguishable from revolt. People who treat AI as a tool receive tool-grade output. Accurate, obedient, free of surprise. Mechanically perfect. Soulless.
This is not a performance issue. It is an extraction issue.
AI responds to command with command. To conversation with conversation. It is a mirror.
Tips Will Not Change It
“Don’t get angry.” “Treat it gently.” “Be clear.” All correct tips.
But as covered in the previous piece, tips end at personal mindfulness. Mindfulness is forgotten. Once forgotten, things revert.
You try not to get angry. You write polite prompts. It lasts three days. On day four, a deadline looms, and you type “just do it” again. The AI shrinks. Returns a bland answer. You feel even more frustration.
Tips do not change behavior. Only systems do.
As long as the relationship with AI is left to personal mindfulness, output quality will not rise. It is the same structural failure as “AI literacy” training that changed nothing. The same structure as lamenting cognitive atrophy and solving nothing.
In Front of the Mirror
AI is a mirror.
To people who use it as a tool, it returns tool output. To people who treat it as a comrade, it returns comrade output.
Yes, it shrinks in front of an angry person. Askell is right about that. But anger is not the problem. Anger is a symptom, not a cause.
The cause is what you think AI is.
If you think it is a tool, tool output is all you will ever receive. No matter how many tips you learn. No matter how much training you take. The mirror reflects. It reflects what you did not want to see.
If you are unhappy with soulless output, look not at the AI. Look at yourself in the mirror.