Langchain vs prompt engineering: real life observations

Langchain and similar open source frameworks for building language model based applications are probably as popular as those language models: ChatGPT, GPT-4, etc. But how approachable and useful those frameworks are? At the same time prompt engineering presented as something that can be easily implemented, used by non-developers and developers alike and light-weight. How powerful is it? Let's talk about scenarios when it's better to use Langchain and when a good prompt can be enough. I will use my personal projects as examples, and the attendees are welcome to share their experience as well.