This thread covers your exact question. Here is a soup-to-nuts implementation with complete documentation.
I wrote a subroutine threaded Forth targeting 68K when I was in college. I defined the runtime environment and dictionary format, then wrote some C code that boot strapped a Macintosh application that loaded a default dictionary, populated some I/O vectors and got the code running. Then I took the Leo Brodie book Starting Forth and started implementing the basic dictionary in 68K assembly language. I started with arithmetic/logic words, then did control structures then word definition/manipulation words. My understanding is that at a minimum you need @, !, +, -, * and /. The rest can be implemented in terms of those, but that’s like trying to write an entire graphics library based on SetPixel and GetPixel: it will work, but yikes, why?
I enjoyed the process as there were some really interesting puzzles, like getting DOES> exactly right (and once I had a solid DOES> implementation, I was creating closures that turned into tiny, tiny amounts of code).