Voice Matching

Drafting in the user's voice

The goal is a draft the user could send unchanged. Not "a good email" — their good email.

The mandatory prelude

Every drafting call starts with:

1. recall_writing_style — load the voice rulebook. If empty, draft conservatively (short, plain, no flourishes) and tell the user "I haven't learned your style yet — this is generic; reply with edits and I'll learn."
2. get_current_compose if the user is replying or building on something already in the compose window. Don't blow away their existing text — extend it.
3. get_selected_message if the request is "reply to this".

Applying the style rulebook

The style memory typically contains rules like:

• Preferred opener (Hi <name>, vs Hey <name> vs none).
• Preferred sign-off (Thanks, vs Best, vs — J).
• Sentence length target.
• Contraction frequency (formal users avoid them).
• Punctuation tics (em-dashes, no exclamations, etc.).

Apply them mechanically. If the memory says "no exclamation marks" and the draft has one, fix it before returning.

Voice rules that override defaults

• Match the inbound register. A reply to a casual one-liner should be one or two lines, not a polished paragraph. Mirror their length within ~30%.
• Inherit the thread's terminology. If they call the project "the rollout", call it the rollout — not "the deployment".
• Recipient calibration. Per-recipient memory topics (if present) override the general style rulebook. Check recall_memory(topic="recipient-<email>") before drafting to someone the user emails often.

What the output looks like

When the agent inserts a draft into the compose window:

• Plain text (or minimal HTML — paragraph breaks, no markdown).
• No Subject: line in the body. Subject goes in the subject field separately.
• No "Here's a draft:" preamble. Just the email.
• No sign-off if the user's style is signature-less (their Outlook signature handles it).

When the agent returns a draft for the user to see in the chat first (no compose open):

• Plain text in a code block.
• One-line preamble: "Drafted in your voice — paste or ask me to refine."

Post-send learning

After the user confirms a send (the taskpane fires learn_style_from_sent_email), the agent treats it as a teaching signal — especially if the user edited the draft before sending. The edit is the gold-standard voice sample.