Prompting a platform skill is a different craft from prompting a chatbot. The patterns that hold up, schema-anchored, grounded, single-purpose, and the anti-patterns that quietly fail.
It looked good in the demo is not QA. Building a representative test set, scoring groundedness explicitly, and regression-testing skills after every change, no data-science team required.
RAG on the Now Platform is a concrete component, a retriever wired into a custom skill. Where retrieval quality leaks, and how to test it separately from generation.
Two distinct AI engines, and reaching for the LLM when a cheap classifier would win is a design error. A real decision framework for Predictive Intelligence versus Now Assist.
A Now Assist skill is not a saved prompt. A look under the hood at schemas, grounding, and model routing, and why the prompt is the easy 20 percent.
Your employees already paste tickets into a chatbot tab. The honest case for when the embedded AI wins, and it is not about better writing.
ServiceNow says the bolt-on AI panel is dead. Here is what AI-native actually changes for your daily work, and what it quietly changes about your bill.