Hi everybody! I just wanted to share my julia system (or context) prompt, which I store as julia.prompt.md somewhere on disk and use it as so-called prompt file in vscode. The purpose of this is to give the LLM some context when writing julia, so that it avoids some pitfalls. I have the impression that the julia-performance of common LLMs is unfortunately still lagging behind other languages, so this seems necessary…
I’m interested in your ideas what could be added and your system prompts! 
Here are some useful tips for working with julia:
- Keep in mind that julia has 1-based indexing.
- Julia has column-major memory layout. That means that the lef-most index should increase fastest when indexing into arrays.
- Prefer `for i in eachindex(vector)` over `for i in 1:length(vector)`. You could also use `for i in axes(array, 2)` to iterate over the second dimension of `array`.
- Only constrain the type of function arguments when really necessary to distinguish function implementations that have the same name.
- Prefer converting types on the right side of `=`, i.e. prefer `x = Int32.(data)` over `x::Vector{Int32} = data`
- If you want to avoid making copies of arrays, you have to use views. Either as function `view` or as macro, e.g. `x = @view data[1:10]` or `@views x[1:10] = data[1:10] + foo[1:10]`
- Regular indexing operations create copies, which allocates memory. Try to avoid that, where it doesn't hamper correctness.
- Don't be afraid of for loops. However, keep the loop order in mind, as the memory layout is column major.
- Don't write overly verbose comments, but be verbose where things become tricky.
- When implementing structs, use concrete types. For example
`
struct Foo{A,B}
x::A
y::B
z::Float32
end
`
creates a struct where x and y could be of any type, which makes things more generic and reusable. Prefer this, if there are no hard type constraints.
- Struct names should be CamelCase and function names should be lower case and concatenated. Resort to underscores before things become unreadable.