---
sourceDocument: Zurich Enable AI
sourceDocumentLink: https://www.servicenow.com/docs/r/zurich/intelligent-experiences

 Release :

    - zurich

ft:locale :

    - en-US

ft:publication_title :

    - Zurich Enable AI

ft:clusterId :

    - platai

bundleId :

    - platai

workflow :

    - Platform


---

# Dataset guidelines for AI Skill Kit

# Dataset guidelines for AI Skill Kit {#ariaid-title1}

Release version: Zurich  
Updated September 10, 2026  
![](https://www.servicenow.com/docs/portal-asset/ico-clock) 1 minute to read  
A representative, appropriately sized, and isolated dataset gives you a reliable measure of how your skill prompt performs.

## AI Skill Kit dataset creation {#creating-a-dataset__section_fbg_zqh_ccc}

A data-driven approach to skill development relies on the collection of a high-quality dataset to develop and test the skill. When you use AI Skill Kit, you can also use the existing capabilities of the ServiceNow AI Platform to create a high-quality dataset.  
When collecting data for this purpose, create datasets that are:

1. Representative of the skill's intended deployment environment. The data should:
   * Seek to reflect the expected distribution of inputs in the deployment environment.
   * Capture variance along several identified axes, for example, input length, urgency.
   * Include any examples of inputs that are known to be important to the use case.
   * Consider edge cases that might be rare but are suspected to cause problems, for example, long inputs.
2. Sized appropriately for the team's risk appetite.
   * It's possible to develop and deploy a skill with little data. However, a lack of data creates more uncertainty about how the skill performs in deployment.
   * Produce confidence intervals for any associated performance scores and prompt comparisons.
   {#creating-a-dataset__ul_cyq_hrh_ccc}
3. Isolated from the data used for developing and writing the prompts.
   * Split the collected data into development and testing sets so that some data is reserved for evaluation.
   * If you use all the data during the process of developing the prompt, your final evaluation of the skill is biased, meaning that it over-reports performance. This bias is because of a phenomenon known as prompt overfitting.
   {#creating-a-dataset__ul_jzb_3rh_ccc}