The Prompt Is Not the Product: What's Actually Inside a Now Assist "Skill"

Every time I explain Now Assist skills, someone says: "Oh, so it's just a saved prompt." And every time, I have to gently break the news. No. A prompt is maybe 20% of it. The other 80% is the part nobody puts in the demo, and it's the part that decides whether your AI is trustworthy or a liability.

Let me open the hood.

A skill, built in the Now Assist Skill Kit, has four real components. The prompt is one. But wrapped around it is an input/output schema, a contract that says exactly what data goes in and exactly what structure comes out. This is the unglamorous hero of the whole system. A free-form prompt gives you free-form chaos; a schema gives you something the rest of your workflow can actually consume without breaking.

The third component is grounding, the data the skill is allowed to look at when it answers. This is your hallucination firewall. A skill grounded in a curated, scoped knowledge source behaves. A skill told to "just answer from everything" will confidently invent a process that doesn't exist. Grounding isn't a setting you flip on; it's a design decision you live or die by.

The fourth is model routing, which brain runs this. ServiceNow's default is NowLLM, its own managed model tuned for workflow tasks. But you can bring your own through the generic LLM connector if you've got a reason. The point is that the model is a swappable component, not the product. That should tell you something about where the actual value lives.

Here's the mental model I want you to walk away with: skills are building blocks, not finished products. One skill does one thing well, classify this, summarize that, extract these fields. The power comes later, when an AI Agent picks up several skills like tools on a workbench and composes them to solve a real problem. If you design skills as modular and single-purpose, your agents get smarter. If you design them as sprawling do-everything mega-prompts, you've built a brittle mess that nobody can debug.

So the next time someone tells you it's "just a prompt," smile, and ask them what their output schema looks like. The silence will be instructive.