Exactly three years ago I was busy writing my diploma thesis - it was about a help system which showed the user directly in the application where and how to perform certain tasks. This type of help is called contextual or context sensitive help.
We had two different types of contextual help: One was bubbles that appeared next to buttons or entry fields, telling the user what to do there. The other was the mouse pointer itself, going into the application, clicking onto certain interface elements or entering text. Both types could visualise whole task sequences, not only single actions.
For my thesis I conducted a user test where the performance after learning a software with a conventional tutorial was compared to the performance after learning with that contextual help system. Of course our expectation was that users better learn a software with our new help system. But why?