Skip to content

PROMPT_ADVISOR.md

Latest commit

 

History

History
87 lines (61 loc) · 5.04 KB

PROMPT_ADVISOR.md

File metadata and controls

87 lines (61 loc) · 5.04 KB

Prompt Advisor

Refine your AI Builder custom prompts

Important

  • Features of Prompt Advisor have been partially incorporated in-product
  • Please see here how to use the features in-product.

Develop effective prompts while learning useful prompt engineering skills with the Prompt advisor tool.
Users can enter a prompt and receive a confidence evaluation with detailed reasoning for the score. Additionally, the tool provides a list of suggested refined prompts implementing various prompt techniques. Users can select from these optimized prompts to iteratively refine and improve their input.

image (1)

Prerequisites

  1. A Microsoft Dataverse Environment with AI Builder prompts enabled in the admin center and Copilot Studio Kit
  2. Permission requirements: System customizer security role.
  3. AI Builder credits: Ensure AI Builder credits are assigned to your environment.
  4. Enable the Power Apps Component Framework Feature : Power Apps Component Framework Feature

Fine tune your prompt

Using the prompt advisor involves an iterative process where you enter a prompt, analyze it to receive a confidence score and feedback, and then refine the prompt based on the provided suggestions. This cycle is repeated until the prompt achieves a high confidence score.

  1. Enter a prompt: Type the prompt you want to analyze. 4 - Copy
  2. Click Analyze: Evaluate your prompt.
  3. View confidence score: Observe the score and feedback on the evaluation pane. 5 - Copy
  4. Review refined prompts:
    • Click the expand button to view suggestions in full screen. 6 - Copy
    • Filter prompts using the technique dropdown menu. 7 - Copy
  5. Select a refined prompt: Choose a suggested prompt to improve your input. 8 - Copy
  6. Click Apply and refine: Iterate based on feedback to improve the confidence score. 9 - Copy Repeat steps 1-6 until you've reached an acceptable confidence score, or the prompt satisfies the needs for your use case. As you learn more about prompt engineering, you may observe it as an iterative process.
  7. Copy the prompt: Paste the refined prompt into the AI Builder custom prompt studio.

Interpreting the confidence score

The confidence score is a numerical value that indicates the likelihood that your prompt will yield accurate and relevant results. Here’s how to interpret it:

  • High (80-100): This indicates that your prompt is well-structured and likely to produce accurate results. The reasoning provided will highlight the strengths of your prompt.
  • Medium (50-79): This suggests that your prompt is good but may need some improvements. The reasoning will point out areas that could be enhanced for better performance.
  • Low (0-49): This indicates that your prompt may not be effective and requires significant changes. The reasoning will provide feedback on what aspects need to be improved.

Tip
If there are too few words entered, the evaluation might not be able to provide a response.

Troubleshooting and FAQ

What should I do if I don't see the Analyze button or get an AI Builder credits error?

Ensure you have the necessary permissions and sufficient AI Builder Credits in your environment.

How can I get more AI Builder Credits?

Purchase additional credits through your Microsoft account or contact your administrator.

Why is my confidence score low?

The confidence score depends on clarity and relevance. Review the evaluation pane feedback to improve the prompt.

What do the different techniques mean?

  • Zero-shot prompting means giving the AI a task with just instructions and no examples, relying on its general knowledge to figure it out.
  • Few-shot prompting means showing the AI a few examples of the task first so it can learn the pattern or style before giving the answer.
  • Chain-of-thought prompting means asking the AI to explain its reasoning step by step, which helps it solve complex problems more accurately.

Back to the landing page