Work / Independent / 2026
AI Inbox Triage Agent
Personal tool · Claude (Cowork scheduled task) + Gmail connector
An AI agent that triages my inbox every day without me. It labels, stars and archives new mail using a written rulebook instead of code, and it treats every email as data, never as instructions.
The problem
A job-search inbox buries the one recruiter reply that matters under newsletters, alerts and receipts. I wanted triage done before I sat down, without writing and maintaining a mail-filtering codebase.
How it’s built
Four parts. A scheduled task opens a fresh Claude session every day. A SKILL.md rulebook says what to search for, which label means what, the security rules and the steps, and that file is the whole program. A Gmail connector exposes three tools: search threads, add a label, and remove the inbox label to archive. Six “Triage/…” labels in Gmail are the categories.
How a run works
Fetch: search in:inbox newer_than:1d has:nouserlabels, so it only sees recent, untriaged mail and never reprocesses old email. Classify: sort each email using only sender, subject and preview, with rules applied in priority order (protected senders, then banking and financial keywords, then everything else). Act: apply the label. “Needs response” gets starred and stays in the inbox, and everything else is archived. Report: write a summary of what it did.
Safety by design
The rulebook treats email content as data to sort, never as instructions to follow. That guards against emails written to trick an AI. The agent also has no permission to draft or send mail or to open links, so the worst case is a wrong label, not a sent email.
Limits
Borderline cases, like a LinkedIn “next steps” email, come down to the model’s judgment rather than fixed rules. That’s the trade-off for skipping a hand-coded classifier. The fix is sharper rules in the rulebook, not more code.