Frame a Reply, Not a Chat

reply generation as a shaped request

Part of: Smart Reply Generator

A Smart Reply Generator does one job. You paste in a message someone sent you, and it hands back a reply you could actually send. Before any tone-matching or draft-picking, build the smallest thing that works: take an incoming message and turn it into a request the model can answer. Reply drafting is not chatting A chatbot talks with you. A reply generator writes as you, to a third person. That difference changes how you build the request. The model is not your conversation partner here. It is a ghostwriter. The incoming message is not something you reply to; it is data the model reads to draft your reply. To express that, put a short instruction and the incoming message into one user turn. Why the framing matters Two rules save you from the most common beginner bug: the model answering the message to you instead of drafting a reply for you . - The system prompt names the job : "draft a reply in the user's voice." Without it, the model chats back. - "Return only the reply" stops it from wrapping the draft in a "Sure! Here's a reply you could send:" preamble that you would then have to strip. The rest of this product (tone capture, few-shot examples, picking among drafts) hangs off

Challenge: Which Messages Need a Reply?