The task
The author started each morning with a full inbox: replies to write, urgent messages to find, and a pile of newsletters, notifications, and receipts to sort or delete. Doing this by hand before the real work of the day took a lot of time.
What they did
The author wrote a plain text rules file with five groups: skip, urgent, needs a reply, for information, and archive. Each group names specific senders and trigger words, because the assistant cannot know who counts as important. One prompt then tells Claude Cowork to read the rules file, open the mailbox, read every unread message, and sort each one by the rules. For messages that need a reply, it drafts an answer from the thread. It saves a one-page summary with the urgent items and the reason for each, the reply drafts, the items for information, and a count of archive candidates. On the first run it archives nothing and puts a review label on each archive candidate, so the author can check the sorting for a week before letting it archive directly.
What the source reports
According to the author, a morning inbox of 127 emails used to take 90 minutes to sort, reply to, and flag. The author says reviewing and approving the summary now takes 12 minutes, which gives back 78 minutes. The author reports about 90 seconds of review per reply, compared with about 5 minutes to write one from scratch. According to the author, sorting accuracy started around 80%, so roughly one in five emails landed in the wrong category, and rose to about 95% after a week of tuning the rules.
Where the approval sits
The assistant only drafts. According to the source, the Gmail connector can create drafts but cannot send, so the author reads each draft in the summary, edits it, and sends it personally. Archive candidates are labeled for the author to check instead of being archived. The source adds that sending through screen control is possible only after the assistant pauses for approval, and that the author keeps a standing rule never to send automatically to an outside recipient.
Worth knowing
The source is one person's newsletter post, so the figures come from a single mailbox and the author's own measurements. The prompt and the sample drafts in the post use placeholder names. The author also describes a mistake: a rule that let the assistant send scheduling confirmations directly produced a reply that confirmed the wrong day after a client asked to move a meeting. The author then ended automatic sending to outside recipients. The source shows no calendar use and no meeting booked by the assistant.
Source: Developer community post · Jun 2026
Drafted with AI assistance and checked automatically. A person is responsible for the collection. How these entries are made
More in this category
- Inbox and scheduling
A web agency sorts its Outlook inbox with AI and approves the drafts
A small web-design agency used an n8n workflow to classify the messages in its Outlook inbox and draft replies for the harder ones. The owner approves each of those drafts with one click, and the workflow logs its actions to a Google Sheet.
Read the entry - Inbox and scheduling
An n8n template labels Gmail threads and saves Gemini reply drafts
An n8n workflow template reads each new Gmail message with a Gemini agent, labels it, and saves a reply draft where one is needed. Each draft waits in Gmail, with the message marked unread, for a person to review.
Read the entry
Ninety minutes live on your tools. One automation running before we hang up.