If you want to perform a robust quantitative analysis, but simply want to do it by hand (because there is less of a learning curve or you want more control in the analysis), you can use the lexicon of automated systems to hand code the data. For example, the lexicon for the VADER sentiment analysis program provides the polarity for many words, and is available here: vader lexicon. I would recommend using the actual tool though.
You could instead use multiple coders and more qualitatively analyze each sentence for polarity, or how positively or negatively weighted an utterance is, and for valence, or how intense the polarity is. As long as your coders have sufficient intercoder reliability, you could make the claim that the coding scheme is robust.
Approaches to sentiment analysis are either automated and highly lexical, or subjective to some degree. I'm not sure if there is a robust middle ground.