Build Your First Now Assist Skill, No Model Training Required

The single biggest myth standing between people and building real AI on ServiceNow is this: "I'd need to know machine learning." You don't. You won't train a model. You won't touch a neural network. You'll compose a skill out of parts that already exist, and if you can think clearly about inputs and outputs, you can do this. Let me walk you through your first one.

Step zero: the right mindset. A skill is a small, single-purpose unit of AI work. "Summarize this incident." "Extract the vendor and amount from this request." Resist the urge to build something that does five things. One skill, one job. You'll thank yourself later.

Step one: prerequisites. You need a ServiceNow instance with Now Assist and the Now Assist Skill Kit (NASK) available, and the appropriate admin role to build. Confirm that before you start so you're not debugging permissions when you should be designing.

Step two: define your inputs and outputs first. This is the step beginners skip and experts never do. Before you write a single word of prompt, decide exactly what data goes in (the incident description, say) and exactly what structure comes out (a three-sentence summary, or a JSON object with vendor and amount). This contract is what lets the rest of your workflow actually use the result. Design it on paper before you touch the builder.

Step three: write and ground the prompt. Now you write the instruction, clear, specific, with an example or two of the output you want. Then ground it: point the skill at the specific knowledge source it should reason from, not "everything." Tight grounding is the difference between reliable and hallucinating. Narrow is your friend.

Step four: pick your model. Default to NowLLM, it's tuned for exactly this kind of workflow task and it's the right call for the overwhelming majority of skills. Only reach for the bring-your-own-LLM connector if you have a concrete reason, and you don't, not on your first build.

Step five: test in the builder. Run real examples, including ugly, messy, real-world inputs, not just the clean one you designed for. This is where you find out your output schema breaks on the edge cases. Good. Better to find it now.

Step six: deploy to Virtual Agent. The fastest way to get your skill in front of real users is to surface it through Virtual Agent, so people in the Employee Center or on mobile can actually reach it.

The first-build mistakes to avoid: making the skill do too much, grounding too broadly, and skipping the output schema. Nail those three and your first skill will feel less like magic and more like what it is, solid engineering you just did yourself.