@scutifer Thanks for the PR. I don't understand its purpose though.

Can you give me more details and maybe an example or two?

@karthink Sure, apologies for the poorly worded PR.

From what I gathered, gptel prefers system prompts as instructions to the language model while user prompts are considered data that the instructions act upon. System prompts are set from gptel--system-message with optional next-message-only suffix from gptel--get-directive.

I wanted the next-message-only suffix feature for user messages too, where I'd add temporary "clarifications" just for that request. This PR is an attempt at duplicating the system directives logic to user messages too. (Models have a different affinity to system vs user prompts; commonly, system prompts are very strongly adhered to while user prompts are soft requests)

• gptel--infix-add-user-instruction is nearly identical to gptel--infix-add-directive
• gptel--get-user-instruction gets the additional suffix to apply during gptel--suffix-send
• the existing code to concat region+instruction is replaced with more explicit normalisation logic that uses your transforms.
• Helpers: gptel--suffix-join-nonempty joins args with 2 newlines; gptel--suffix-normalize-prompt uses join-nonempty to handle both cons cells and strings. gptel--suffix-insert-transform adds the given transform before gptel--transform-add-context. gptel--suffix-make-instruction-transform returns a transform to append instruction at send time. These exist so that we can handle 3 cases
  1. explicit prompt, instruction exists: instruction is merged immediately to the prompt
  2. implicit prompt from buffer, instruction exists: a transform is created to append instruction when sending it to the LLM
  3. explicit prompt, no instruction: use the prompt as-is, no transforms

(tbh there maybe a simpler solution but I couldn't think of any as I'm not very familiar with the codebase)