What they did
Abdallah Deeb works on a small infrastructure team and spends much of the day sending quick asks like a review request or a staging check. He already runs Claude Code most of the day, so he wrote a skill plus a small Python script, mattermost-dm, that lets the agent send Mattermost DMs as him. Before drafting, the agent reads a private voice.md style guide. Rather than write that guide by hand, he had Claude Code pull 90 days of his DM history with a few colleagues and draft it from only his own messages, with no real names or quotes. The script adds a greeting only when he has not spoken to that person for a few hours, and fills first names in batch messages. Every message goes through a dry-run preview, and only the exact previewed text is sent once he says yes.
How it went
Before the style guide, the agent turned one-line questions into five paragraphs with bullet points and a sign-off, and colleagues left them for hours. With the guide, drafts read like the short notes he would type himself, and he tunes the file until test messages need no edits. He reports no counts of messages sent or time saved. He published the skill as MIT-licensed code.
Worth knowing
Pulling DM history sends both sides of each conversation, including colleagues' messages, to the AI provider while the guide is drafted, so he advises checking company policy or writing the guide by hand. The list of things you never write did the most to stop padded, AI-sounding messages. The access token stays in a file the agent is told not to open.