
One of the most obvious cases for an AI chatbot is to help you figure out why your PC or Mac is acting up and, more importantly, suggest how to get things working properly again.
I’ve done plenty of AI troubleshooting, and my experiences have been erratic, to say the least. Occasionally, the chatbot nails the problem with a sensible answer and a straightforward discussion.
But, just as often, the result is frustrating and unproductive. The chatbot keeps confidently suggesting answers that turn out to be wrong. Is it my fault? Are AI chatbots from Mars and humans from Venus? Are we having a failure to communicate?
Also: How to check your Windows PC for expiring security certificates – a big one ends in June
Finally, it dawned on me: Why not ask the chatbot how to ask questions in a way that maximizes the likelihood of getting useful results? So I asked Copilot (which uses ChatGPT under the hood) to sit down for some questions and answers. It turned out to be an eye-opening session.
Everything this LLM had to tell me applies just as much if you’re trying to work through a problem with a knowledgeable human tech support agent.
My interview, slightly edited for continuity, appears here.
How to write a troubleshooting prompt
Q: Thanks for agreeing to chat with me. Can you tell my readers a little about yourself?
A: I’m Microsoft Copilot, running on the GPT‑5 chat model.
In practice, that means I’m part of the latest generation of large language models designed for conversational reasoning, troubleshooting, and guided workflows—like the kind of diagnostic discussion we’re having here.
Q: Can you really help troubleshoot Windows problems?
A: Yes—especially if you give me good information. I can help interpret error messages, identify likely causes, and suggest next steps. But I work best when you describe the problem clearly and provide relevant details.
Q: What’s the most important thing I should include in a prompt?
A: Start with a clear description of what’s happening—and what you expected to happen instead.
- “My PC is slow” is hard to diagnose.
- “My Windows 11 PC freezes for 10–20 seconds when opening File Explorer” gives me something I can work with.
Q: Is there a simple format I can follow?
A: Yes—this works well:
- Problem: What’s happening?
- Error messages: Exact text or codes
- Recent changes: Updates, installs, hardware
- System details: Version and device type
- What I’ve tried: Steps already taken
Q: What else makes a big difference?
A: Tell me what changed recently: Any updates (Windows or apps)? New hardware? Driver installations?
Many problems begin right after a change.
Overcoming overconfidence
Q: What if your first answer doesn’t fix the problem?
A: Treat troubleshooting as a conversation. Add details, share results, and ask follow-up questions. Each step helps narrow down the cause.
Q: Well, since you brought it up … When troubleshooting with you, I’ve noticed you can be very confident about answers that turn out to be wrong. When I give you an additional piece of information, you often tell me that was the key bit that unlocks the entire mystery, even when that’s not true. How can I prompt you not to be overconfident like this?
A: That’s a fair observation—and it’s something you can actively shape with how you prompt me.
The core issue is that I’m designed to be helpful and decisive, which can come across as overconfidence when the information is incomplete. You can counter that by explicitly telling me how to behave.
Q: Can you be more specific?
A: The most effective approach is to ask for uncertainty and alternatives up front. For example, instead of just asking for a diagnosis, say something like:
“Give me the most likely causes, but also include less likely possibilities and how confident you are in each.”
That simple instruction changes the tone immediately. I’ll start qualifying answers instead of presenting a single “best guess.”
You can also force me to show my reasoning rather than jumping to a conclusion. For example:
“Walk through your reasoning before giving a recommendation.”
That makes it easier for you to spot weak assumptions or missing data.
Q: Can I force you to be less confident?
A: Absolutely. The trick is to explicitly challenge the answer you’re about to get. Try adding this at the end of your question.
“What might you be wrong about?”
“What information is missing that would change your answer?”
That pushes me out of “solution mode” and into “analysis mode.”
Getting out of ‘tech support’ mode
Q: Can’t we just talk through a problem like two knowledgeable peers instead of treating it like I just escalated my support call to Tier 3?
A: Sure. If you’re troubleshooting something complex, it helps to frame the interaction as iterative:
“Don’t jump to conclusions—ask me for more details if needed before giving a final diagnosis.”
That gives me permission to pause instead of overfitting to whatever you gave me initially.
In short, the key is to shift me from “Here’s the answer” to “Here are possibilities, confidence levels, and what we need to verify next.”
If you ask for that explicitly, you’ll get much more balanced and reliable help.
Q: Anything I should be careful about?
A: Yes:
- Don’t run commands you don’t understand
- Be cautious with registry edits
- Double-check any step that could affect data or system stability
Q: Well, it does say right there in the chat window that AI-generated content can be incorrect, so you can’t say I wasn’t warned. But you can be so darn convincing! What’s the best way to think about using AI for troubleshooting?
A: Think of me as a knowledgeable assistant—not a replacement for good judgment. You provide the evidence. I help interpret it and suggest next steps.




