AI Prompts to Triage a Full Inbox in One Sitting

The triage problem is easy to describe and painful to live with. You have two hundred unread emails, no clear idea which five need a response today, and two bad options: open each one and spend the morning reading, or skip the inbox and risk missing whatever is genuinely urgent. The right set of ai prompts to triage your inbox solves this — it classifies every thread in a single pass and hands you an actionable shortlist instead of a pile.

This guide is that prompt set. Below you will find a bucket-definition prompt that teaches the model your classification system, a one-line-per-thread scan that processes your whole inbox at once, a filter that surfaces only what genuinely needs you today, and a defer-and-delegate pair for the rest. The prompts work in ChatGPT, Claude, Gemini, and Microsoft Copilot — and in any AI assistant with direct mailbox access, which removes the copy-paste step entirely.

What do you need before running an inbox triage prompt?#

Two things have to be in place before the prompts can produce a useful result.

The first is a set of triage buckets that match how you actually deal with email. The prompts below default to four: Reply now (needs your response today or tomorrow), Read and decide (worth reading, no urgent reply needed), Delegate (needs action but not from you), and Archive (newsletters, notifications, receipts, and CC threads with no direct ask). These defaults work for most inboxes, but they sharpen considerably when you tune them to your role. A recruiter might add a Candidates to advance bucket. A founder might split Reply now into Investor-priority and Other. Spend two minutes naming your actual buckets before you start — the model applies them exactly as written, so sharper definitions produce sharper sorts.

The second is what you paste. If your AI assistant cannot connect to your mailbox, you are working from a text export. The most useful format is sender, subject line, and the first twenty to thirty words of the preview, one thread per line, for every unread message. In Gmail, switching to compact view and copying the visible list works well. In Outlook, the message list pane gives you the same data. You do not need full email bodies for triage — subject lines and snippets are enough to classify most threads — and keeping body text out of the paste also limits how much real correspondence you are sending to a third-party server.

How do you define your triage buckets for the AI?#

Step one is a system-level instruction that defines your buckets in one message. Paste this at the very start of the session, before anything else. It primes the model for every prompt that follows, so you do not have to re-explain your classification rules inside every scan.

The ESCALATE bucket is doing quiet but important work here. It gives the model a safe exit for threads it cannot confidently classify from a subject line and snippet alone. A wrong REPLY NOW costs you time; an honest ESCALATE costs you nothing — you review it manually, the same as you would have without the triage.

What is the prompt to scan your whole inbox in one pass?#

Once the model has your bucket definitions, the scan prompt does the work. Paste your sender-subject-snippet list under this prompt. The output is a table — every thread classified in a single pass, with a one-phrase reason so you can spot any disagreements at a glance.

The sort order matters. Asking for REPLY NOW first and ARCHIVE last means the threads that need you appear at the top of the response, before you have scrolled through the noise. Run this in the same session immediately after Prompt 1 — the model already has the bucket definitions in context.

How do you extract what genuinely needs a reply today?#

The scan gives you a sorted table. Prompt 3 narrows it to the threads that actually need you, ranked by consequence, with a suggested first move for each. Run it in the same session immediately after the scan — the model already has the full triage context.

The TRUE URGENT rule in the third line is worth keeping even when your REPLY NOW list is short. When a large inbox produces eight or ten REPLY NOW threads, re-ranking by consequence separates the five highest-stakes items from the rest, which prevents the list from feeling equally urgent and stops you from starting at the wrong end.

For the DELEGATE threads, a fourth prompt drafts the forwarding message for each one. These are two sentences each — direct, written to a capable colleague, with no explanation of why you are delegating. The model does not need to know who the recipient is; it uses a placeholder you fill in before you send.

Does this prompt work differently in ChatGPT, Claude, Gemini, and Copilot?#

The bucket and scan prompts above work across all the major chat assistants, but each platform has quirks that affect how reliably the triage goes. The table below maps the differences so you can pick the right tool for the job or adjust the prompt when the default output does not behave as expected.

One limitation applies to every platform in that table: chat assistants do not store your inbox between sessions. Every new conversation starts from zero, which means the paste-and-classify loop resets each morning. You can process a backlog in one sitting, but ongoing triage requires a fresh export each time. That structural ceiling is the reason a connected assistant outperforms a chatbot for recurring triage — not because the prompts are sharper, but because the information does not have to move.

What do you do when the triage prompt gives bad results?#

Four failure modes come up often, and each has a direct fix.

The model files everything in ARCHIVE. This usually means the ARCHIVE definition is too loose or the snippets are too short to classify. Tighten the rule — add 'CC only with no specific ask directed at me' so the model is not tempted to file anything ambiguous there. Also paste longer snippets, fifty words rather than twenty. When the model cannot see a clear ask, it defaults to the lowest-friction bucket.

The model puts too many threads in REPLY NOW. Usually caused by a REPLY NOW definition that is too broad. Add an explicit constraint: 'REPLY NOW means there is a specific question or request directed at me, with a real consequence if I miss it today. Do not use it for FYIs, newsletters, or threads where I am CC'd.' That one rule cuts most false positives.

The output loses its table format partway through. A context-length problem. Break the inbox list into chunks of twenty to twenty-five threads and run the scan in two passes, then merge the sorted rows. Adding 'output only the table with exactly these columns: Bucket, From, Subject, Why — no other text' also locks the structure across a longer paste.

The model invents urgency that is not in the snippet. A dramatic subject line can trigger REPLY NOW even when there is no actual ask. Reinforce the ESCALATE rule: 'if you cannot see a specific request or deadline in the snippet I gave you, use ESCALATE.' An honest ESCALATE is always better than a confident wrong classification.

Is there a faster way that does not require pasting anything?#

The paste loop has a ceiling. It cannot process emails that arrive after you started the session, it resets every morning, and it multiplies across multiple accounts. When the manual prompt set starts to feel like its own kind of inbox problem, that is the signal to move to an assistant that connects to your mailbox directly.

We build AI Emaily, an AI-native email client that runs triage automatically inside your real inbox across Gmail, Outlook, and IMAP accounts. You configure your priority rules and Context Brain once; the assistant applies them continuously, surfacing threads that need you and routing the noise without a fresh export each day. Copilot mode queues a suggested action per thread and waits for your one-tap approval before anything is sent. Autopilot mode can archive and label within the rules you define, with an undo and a full audit trail on every action. There is a 7-day free trial on every paid plan — see AI Emaily pricing to compare what is included.