---
title: "Why your AI refuses in one app and not in another"
description: "The same question, the same model, and yet a different answer? Then the app is deciding too. Here is how you read a refusal and know what to do with it."
language: en
source: https://coachsteff.live/en/de-app-beslist-mee
---

# Why your AI refuses in one app and not in another

You ask your AI something very ordinary and you get a polite refusal. You try the same thing in another tool, with the same model underneath, and there it works. That actually feels a bit random. In a vibe coding training, a participant ran into this too. She literally said: 'It is really strange. How is that possible?'

The answer is not in your prompt and not in the model either. It is in the app around it. And once you see that, you will read a refusal very differently. You then know where to look. And you also know when it is better to just stop trying.

Copy the prompt at the bottom of this article. With that, you let your AI explain for itself where a refusal comes from.

## What your model receives

A language model receives a whole stack of instructions with every question. Your own question is only a small part of that. On top of that come the instructions for your agent, your project folder, the memory, your skills and your connectors. And right at the top are the system instructions of the app itself. You cannot read those and you cannot adjust them either.

![Around your own prompt are three more layers of instructions, with the system prompt of the app at the top.](https://coachsteff.live/media/929e26921cf1ae7f.webp)

In the workshop I always show that with a drawing. Your prompt is then a small purple block at the bottom. Almost all your attention goes there, and yet it is actually the smallest part of what the model reads. I find that a bit ironic, because we spend hours refining that one block.

The providers write those system instructions themselves. The big providers put their rules in there. Those are about copyright and privacy, but also about the terms of other platforms and about the EU AI Act. If you use Copilot or the app of ChatGPT or Claude, then you are in their armour. And then they determine what is allowed.

Another participant once asked Copilot for a picture of a famous cartoon character. Copilot politely refused, because that is protected material. The model could draw it perfectly. The app just did not allow it.

## Frequently asked questions

**Why does my AI refuse something that works in another tool?**

- Every app sends its own system instructions to the model. These contain rules about copyright, privacy and the terms of other platforms. The same model can therefore behave differently in another app.

**Can I read or adjust those system instructions?**

- No, the provider of the app writes those themselves. In your own environment you largely determine the rules yourself, but then you also bear the responsibility yourself.

**Does it make sense to rephrase a refused question?**

- Only if the cause lies in your own instructions or settings. If the refusal comes from a rule of the app, then the no remains. First ask your AI where the limitation comes from.

**Where do I learn how those layers around a model work?**

- In 'Claude AI Power User' and 'Vibe Coding with AI' you work hands-on with instructions, project folders and your own environments. You can find all dates at [Workshops & Events](https://coachsteff.live/agenda).