An important aspect of building natural-language generation (NLG) systems is developing specific schemas, lexicalisation rules, grammar rules, etc for a particular application. To date this process has mostly been based on corpus analysis. In this talk I will discuss attempts at using some of the Knowledge Acquisition (KA) techniques developed in the expert system community, such as sorting and think-aloud protocols, to acquire the application-dependent knowledge needed by an NLG system; certainly our experience is that combining these techniques with corpus analysis produces much better results than corpus analysis on its own. This work is being carried out in the context of the STOP project, which aims to build an NLG system which generates personalised smoking-cessation letters.