Power Platform Community Forum Thread Details

I have built an agent using the Quiz Tutor template in Copilot Studio Lite. When I upload a source file for generating quiz questions, the agent shows the question and the multiple-choice options correctly. However, it also creates action buttons such as Explain the question and Next question, and it sometimes adds an extra button that reveals the correct answer.

I have already included explicit instructions telling the agent not to do this, including the line: “When generating a question, do not reveal which option is correct in any way, including in text, hints, tooltips, buttons, or suggested actions.” Despite this, the agent still generates a button such as “Answer B”, which exposes the correct option before the user responds. The screenshot shows an example where the question asks for option A, but a button displaying “Answer B” appears underneath.

How can I stop Copilot Studio Lite from leaking the correct answer like this?

2026-01-15_12-14-...

Your file is currently under scan for potential threats. Please wait while we review it for any viruses or malicious content.

Categories:

Copilot Studio pre-built agents/templates

Hi there,

This is a classic case of suggested-actions / follow-up button behavior overriding your written instructions in Copilot Studio Lite. Negative instructions alone (e.g., "do not reveal…") are often not enough because the model still treats the correct answer as in-context and generates suggested actions from it. Here are the official Microsoft Learn–grounded strategies:

1. Rewrite your instructions as positive, behavior-driven rules

Per Write agent instructions – Microsoft Learn, agents respond best to positive, action-oriented instructions in simple, direct language. Replace "do not reveal" with:

"When you ask a quiz question, only produce: (1) the question text, (2) the four answer options labeled A–D, and (3) a single follow-up action 'Submit my answer'. Never produce buttons, suggested replies, hints, or text containing the letter or content of the correct answer until the user has submitted their answer."

Positive constraints with an explicit allow-list (what you CAN produce) work far more reliably than blanket denial.

2. Separate the correct answer from the model's response context

The root cause is that the correct answer is in the same context window the model uses to generate suggested actions. Mitigation patterns:

Don't pass the correct answer to the response-generation step. Have the agent generate only the question + options in one step, store the correct answer in a variable/topic state, and only reveal it in a separate downstream node after the user submits their answer.
Move the quiz logic into a topic/agent flow. Generative orchestration (Orchestrate agent behavior with generative AI – Microsoft Learn) is powerful but probabilistic. For quiz integrity, a deterministic topic flow (Question node → Condition node comparing user input to stored answer → response) keeps the answer out of the response generation entirely.

3. Disable / control suggested actions

In Copilot Studio Lite, the suggested-actions / follow-up buttons are emitted by generative orchestration based on context. You can:

In the agent's Generative AI settings, turn off automatic follow-up suggestions if your channel exposes that toggle.
In your channel (e.g., Microsoft 365 Copilot), confirm that your client honors a setting to suppress suggested actions, since per-channel rendering varies.

4. Don't fight the citation/system formatting

The docs explicitly warn: "Don't modify, override, or interfere with the system-defined citation format or behavior. Avoid instructions that attempt to alter how citations are generated, structured, or displayed." If your prompt is trying to suppress system-generated UI, prefer the structural separation in step 2 instead.

Recommended approach (most reliable):

Move from a single generative "Quiz Tutor" agent prompt to a small topic flow:

Topic generates question + options + stores correctAnswer in a topic variable (NOT echoed).
Question node captures the user's choice.
Condition node checks userChoice == correctAnswer and replies correct/incorrect.
After feedback, optionally call generative AI for an explanation.

This eliminates the leak entirely because the correct answer never sits in the same generative response that produces the UI buttons.

Hope this helps you ship a quiz that doesn't spoil itself!

Raghav Mishra — LinkedIn | PowerAI Labs

Found this helpful? Please mark ✅ "Does this answer your question?" so others searching for the same issue can find it quickly. A 👍 on "Was this reply helpful?" or a ♥ Like is also much appreciated!