Humanizing iMessage assistants. Flow, timing, and voice

Write like a person. Respond at the right tempo. Never spam. These rules sound obvious. They also separate a tool that people keep using from one they mute.

Message flow and timing

• Merge quick back to back messages from the same person for about two seconds, then answer once. This prevents double replies and reads like a single turn.
• When you merge, keep any attachments and combine the text in order. Then treat it as one turn.
• Process once per merged turn. Store. Think. Act. Reply. Do not send two messages for one burst.

How we wire it in code

• We keep a short buffer per sender. Every new message resets a two second timer. When the window closes we merge and process one message.
• Files: rabbit_worker.py keeps per sender buffers and the timer. _process_user_messages() merges and calls the agent. _combine_messages() joins content and keeps unique attachments.
• Note: _combine_messages() uses a set to dedupe content. If you need strict order, switch to an order preserving method.

When not to reply

• Give the model an escape hatch. If it outputs a sentinel such as NO_RESPONSE, send nothing and mark the turn as handled.
• Good uses: one word acknowledgments that add no value. Waiting on a tool. A stray ping with no question.

How we wire it in code

• The policy tells the model that it may return NO_RESPONSE. We check the string before sending and drop the turn.
• Files: services/prompt_builder.py:get_policy_snippet() mentions NO_RESPONSE.
• Where to gate: right after the agent response in rabbit_worker.py:_process_combined_message() or inside unified_sender.py:send() before dispatch.

How the assistant speaks

• iMessage style: be concise. Skip markdown. Be personable and brief. Avoid emoji spam.
• Email style: write a clear subject. Use clean HTML or structured text. Reply in the existing thread when it makes sense.
• Do not leak your thinking steps. Think and act. Then send a single message that reads like a human message.