LLMs may seem like magic buy they aren't. They operate within the confines of the context they're given. The more abstract the context, the more abstract the results.
I expect to need to give a model at least as much context as a decent intern would require.
Often asking the model "what information could I provide to help you produce better code" and then providing said information leads to vastly improved responses. Claude 3.7 sonnet in Cline is fairly decent at asking for this itself in plan mode.
More and more I find that context engineering is the most important aspect of prompt engineering.
LLMs may seem like magic buy they aren't. They operate within the confines of the context they're given. The more abstract the context, the more abstract the results.
I expect to need to give a model at least as much context as a decent intern would require.
Often asking the model "what information could I provide to help you produce better code" and then providing said information leads to vastly improved responses. Claude 3.7 sonnet in Cline is fairly decent at asking for this itself in plan mode.
More and more I find that context engineering is the most important aspect of prompt engineering.