Capture the User's Tone

building a tone profile from sample replies

Part of: Smart Reply Generator

A generic reply is useless. "Thank you for your message. I would be delighted to attend." is technically correct and sounds nothing like how you text a friend. Making the draft sound like you is the core of this product, and it starts by measuring how you actually write. Tone is a set of measurable habits You can't hand the model a vibe, but you can hand it observable habits pulled from your past replies: - Length : do you write 4-word replies or 40-word ones? - Greeting : do you open with "Hey" / "Hi", or dive straight in? - Sign-off : do you close with "Thanks" / "Cheers", or just stop? - Energy : exclamation marks and emoji, or flat and dry? Compute those from a handful of sample replies and you have a tone profile : a small dict that captures your writing fingerprint. You then feed it to the model as explicit instructions: Why measure instead of guess Write "casual and friendly" in the prompt and the model picks its own idea of casual, which drifts toward chirpy customer-service English. Numbers leave no room for that. "About 6 words" is a target the model can hit; "be concise" is not. It is the same lesson as prompt-writing everywhere: specific beats vague, and measured beats

Challenge: Tone Fingerprint