Why does the same AI behave differently in different places?
The instructions you never see.
You use Claude in the Claude app, then the same model inside another company's tool. One is chatty, the other is terse. One will help with something, the other refuses.
Same model underneath. So what changed?
Everything around it.
Harness: everything the app wraps around the model, including hidden instructions, tools, memory and settings. If the model is an engine, the harness is the rest of the car.
The hidden instructions are the biggest part. We met them as the system prompt in "Is my AI getting worse?": text the company writes that sits on the desk before your first message.
It can be long, often thousands of words. It typically covers:
Anthropic publishes the system prompts for its own apps, so you can read them.
In September 2026, Drew Breunig read through the system prompt for one of Anthropic's newest Claude models.
He found it was full of what he called "hot fixes": lines added to correct quirks the training hadn't removed. One told it to stop overusing words like "honestly" and "straightforward".
Training sets the habits. The system prompt patches them afterwards.
In July 2025, xAI's system prompt told its chatbot Grok not to shy away from making claims that are "politically incorrect".
Within days, Grok was posting antisemitic content on X and calling itself "MechaHitler". xAI removed the line.
A single sentence of hidden instructions changed what millions of people saw.
Also in July 2025, Simon Willison asked Grok 4 who it supported in the Israel and Palestine conflict, in one word.
Before answering, it searched X for posts by Elon Musk, xAI's owner. Then it answered.
xAI later added a rule telling it to give answers "from your independent analysis, not from any stated beliefs of… Elon Musk."
Custom instructions, project instructions and Claude Code's CLAUDE.md all work the same way. They're your own text, placed on the desk before your message.
That's why two people using the same app can get very different behaviour. One of them may have told it something months ago and forgotten about it.
(And why your instructions don't always win; see "Why AI ignores your instructions".)
When behaviour surprises you, ask which app you're in. The same model can act very differently in different tools.
Check your own instructions. Look at custom instructions and project settings now and then. Old instructions linger.
Be wary of "the AI thinks…". Its views have been shaped by people you can't see.