system vs user vs assistant roles in the make openai module, what goes where?

agree with all that. one habit that saved me a lot of debugging: keep the system message completely static, no mapped variables in it at all. the moment you build it from data, every run has slightly different instructions and you can't tell whether a weird output came from the input or the rules.

system = the standing instructions. who the model is, tone, format you want back, rules like "return only json". this stays the same every run. user = the actual input for that run, so this is where you map the data coming in from the earlier modules. assistant = what the model said before. you only need it for two things: giving an example answer so it copies the format (put a sample user message, then a sample assistant reply, then the real user message), or passing back earlier replies if you're building a conversation. most scenarios only need system + user. if the output format keeps drifting, the example pair is what fixes it.