What it does
Inbox Triage takes a batch of emails and tells you which ones actually need you. Each message is sorted into urgent, needs action, information only, noise, or suspicious, then prioritised — and senders you name as VIPs are always pushed to the top regardless of what the message looks like. For the messages that warrant a reply, it drafts one. Not sent, not queued: drafted, for you to read and decide on. The classification and prioritisation are fully deterministic, so the same inbox always sorts the same way. Only the reply drafting uses a language model, and even that falls back to a template when none is available, so the triage never fails for want of a model. The guarantee worth reading is what it will not do. It classifies and drafts, and there is no code path that sends, archives or modifies your mailbox. Nothing you hand it can be altered by it. Anything that looks like phishing or spam is flagged and deliberately gets no draft reply — because the one thing worse than a phishing email in your inbox is a polite auto-drafted response to it sitting in your outbox. Email content is treated as data rather than instruction throughout. That matters more here than almost anywhere else: an email is text a stranger wrote and sent you specifically, which makes an inbox the most obvious place in the world to hide an instruction aimed at an assistant. Anything embedded is neutralised rather than followed, and personal information is kept out of the logs. `classify` runs the sorting alone when you just want the split and the priorities without any drafting. `run_full` does the whole pass. No mailbox connection and no credentials — you provide the batch, and nothing leaves. Built for anyone whose morning starts with two hundred unread messages and no idea which three matter.
Sorts your inbox into urgent, action, info, noise and suspicious — and drafts replies it can never send.
Example prompts
- Triage these 10 emails with run_full, treat boss@company.com as a VIP, and draft concise professional replies where needed.
- Run classify on this batch — no drafts, just give me the urgent/action/info/noise/suspicious split and priorities.
- Which of these messages are suspicious? Flag them and don't draft replies for those.