When I first started testing GenAI, I asked it to grade something from 1 to 10. The responses I got included 2, 3, 5… and even 15!? That one really threw me off. Turns out, the issue was that I wasn't being clear enough in my prompt.
When ServiceNow released its Prompting Guidance, it all made sense to me—especially after my own missteps trying to get GenAI to respond the way I expected. Below are ServiceNow’s principles for prompt guidance:
The LLM is smart, but it is NOT a mind reader - learn to ‘speak’ the language to get better results:
- Write clear, specific, and concise prompt
- Specify triggers and conditions
- Specify the expected outcome and context
- Include necessary details like task requirements, specific function names, and any constraints
- Specify timeframes and date ranges
- Prompt for one component part at a time
- Always validate the response returned by the LLM is correct
Weak vs. Strong prompts
More examples…
Weak | Strong |
// close child records when the process ends | // close all child Task records when the parent Task is closed |
// get the count of user cases | // get the count of currently open Cases for the current user with an AJAX call |
// create a related list to show the records opened by this user
| // create a related list on the Incident record to show all Incidents opened by the same caller
|
// pull all the related customer information when an Incident is created
| // create a JavaScript function that uses a GlideRecord query to build an array of Customer sys_id values, and call the function from the condition builder in the Incident table
|
// replace social security numbers with #
| //check the short description and description in the current record for any social security numbers, and replace the SSN information with '###'
|
//check for open cases that have not been updated in 3 months
| //check for open HR cases that have not been updated in 3 months
|
The bottom line, with a little bit of trial and error, you will learn how to write your strong prompts.
Thanks for reading.
