Zum Inhalt springen

12 Terminal Windows: How My AI Learned Orientation

12 terminal windows. A prompt in the wrong one.

This is how I taught my AI to stop me.

Earlier this year I was running several projects in parallel, each in its own terminal window. AI agent on the left, knowledge system on the right, research project in the middle. Up to 12 windows open at once.

It worked. Until it did not.

Eventually, notes from project A ended up in the database for project B. Research findings were saved in the wrong context. Facts got mixed up. The worst part was that I only noticed days later. The AI raised no error; it simply did what I told it to do, in the wrong window.

The problem was not the prompt. The prompt was good. The problem was that the AI did not know where it was.

AI Has No Sense of Space

ChatGPT, Claude, Copilot. They all answer what you ask. None of them ever asks back: Does this question even belong here? Is this the right context? Does this fit what we are currently working on?

A human would notice. If you suddenly ask about recipes in a budget meeting, someone asks if everything is all right. An AI just carries on.

The Solution: Scope Checks

So I taught the system to check its context. Every terminal window gets a scope: a project context that defines which data belongs to it. When I tell the AI agent “Save this note”, the system checks whether the content matches the active scope. If I dictate a note about a different project while in the research project, the agent stops and asks: “This doesn’t sound like the project we’re working in right now. Right window?”

It sounds simple. It was. A few rules, a comparison, a warning. No machine learning, no fine-tuning. Just the question: Does this belong here?

Yet almost no AI tool asks this question on its own.

The Real Insight

We talk a lot about better prompts: prompt engineering, context windows, system instructions. Almost nobody talks about what happens when a perfect prompt lands in the wrong context. The more windows, projects and agents you run in parallel, the more critical that becomes.

It is not the quality of the answer that matters. It is whether it lands in the right place.

The right prompt in the wrong terminal window.