Cross-validation value

https://app.dataquest.io/m/185/getting-started-with-kaggle/9/making-predictions-on-unseen-data

What actually happened: 
Actually I get confused of the process : 
First : we instantiate the model & working on it 
second : we used cross validation  & instantiate another model & working on it 
Third : after check our model accuracy with the accuracy of cross validation we instantiate another model like the first one & used it on final prediction 
Now I have the following question : 
1- should I follow these process on any model , I mean instantiate the model first then use cross validation ?
2- cross validation model is only useful on comparing accuracy & to make final prediction should I use the original model without cross validation?
3- What if the accuracy of cross validation was far than the original model what should I do ?