Inbox Triage Agent Specification With Human Approval Points
Specify an inbox triage agent that sorts email, drafts replies and asks for human approval before it sends anything.
Added October 15, 2026. Written with AI assistance and checked for clarity and placeholders; not reviewed by a lawyer, accountant, clinician or other professional. How these prompts are made
The Open buttons send the prompt text as currently shown below to that site only when you click. Fill in the placeholders first.
[VERIFY] is an editorial flag, not a blank to fill. Wherever the prompt shows [VERIFY], check that item against a reliable source before you use it (what [VERIFY] means).
Write a specification for an agent that triages my email inbox. My role and the kinds of email I get: [ROLE_AND_EMAIL_TYPES]. Categories I want (for example needs reply today, waiting on others, read later, newsletters, finance, personal): [CATEGORIES]. Tools the agent can use (read mail, apply labels, draft replies, create calendar holds): [AVAILABLE_TOOLS]. Things it must never do without asking me: [APPROVAL_REQUIRED_ACTIONS]. Return: 1) a one-paragraph purpose statement; 2) a table with Category, How the agent recognises it (sender, keywords, thread state), Action it may take alone, Action that needs my approval; 3) a step-by-step run: fetch, classify, explain the classification in one line, draft only for categories I allow, queue for approval; 4) the exact approval message format (what it shows me, with Approve, Edit, Skip); 5) a list of twelve test emails I should try, including ambiguous ones, a phishing-style message and one with sensitive content; 6) the failure behaviours: what it does when it is unsure, when a tool fails, and when two rules conflict. Draft the system instruction for the agent in under 250 words in plain language. Mark any capability I did not list as [VERIFY] rather than assuming it exists.
Fill in before use
Optional: type your details and the prompt above updates. Nothing is saved or sent.
Replace each bracketed placeholder with your own details, then paste into Any AI or another assistant.
See an example result (fictional data)
Category | Recognised by | Agent alone | Needs approval Needs reply today | direct question to me, sender in my team list | label, summarise in one line | any draft reply Newsletters | List-Unsubscribe header, sender not in contacts | label, move to Read later | unsubscribe Finance | words invoice, payment, bank | label, flag | forward or reply Approval message: 'From Dana Lee: asks if Thursday 3 pm works. Draft: Thursday 3 pm works, I'll send an invite. [Approve] [Edit] [Skip]' If unsure: label 'Review' and say why in one line; never guess a sender.
Illustrative only, made-up data. Your results will differ and must be reviewed.