The email triage playbook

The inbox is the most reliable distraction in modern work. You open it to check one thing, you close it forty minutes later, and somehow the actual reason you sat down has evaporated. "Inbox zero" is a religion with a thousand denominations. The real goal is simpler: the inbox shouldn't run your morning. AI gets you there, cheaply, in an afternoon of setup.

Here's the playbook.

Four buckets, every email

Almost every message you get falls into one of four categories. Get this taxonomy right and the rest is plumbing.

VIP. Your boss, your top three customers, the investor who texted last week, the partner whose deal closes Friday. Surface immediately. Push notification, top of the inbox, never auto-reply, never auto-archive. AI doesn't touch these.

Action required. A real person needs a real response or a decision. AI drafts the reply, you approve. This is the bucket where AI buys back the most hours.

FYI. Newsletters, weekly project updates, status pings, calendar confirmations you've already seen. Nothing for you to do. AI summarizes and archives. You read the digest once a day.

Noise. Cold sales pitches, "just checking in," LinkedIn invitations, no-ops, third-party re-confirmations. Auto-archive. You'll never miss them.

If you can't put an email in one of those four buckets, your buckets are wrong.

Three setup paths

Pick by how much you want to build versus how much you want to install.

Easiest. Lindy, or any of the AI agent platforms. Drag-and-drop rules, point it at your inbox, it auto-categorizes and drafts replies in the background. Superhuman now ships AI categorization and drafting baked in if you want it done for you and don't mind the price tag. Both work. Neither requires code.

Middle. Zapier or Make with AI Actions. Trigger on inbound mail, pass the body to Claude or GPT-4 with a classification prompt, route the output to a Gmail label, draft a reply, or ping you in Slack for VIPs. Twenty minutes of clicking, no scripts.

Custom. A 50-line script hitting the Gmail API on a cron, piping each new message into your own model call, writing labels back. Pick this if you already write code and want full control over the prompt, the logging, and where drafts land. Honestly the most fun, also the most maintenance.

The classification prompt

Whichever path you pick, the model needs a real prompt. Few-shot examples beat any "be a helpful email assistant" opener.

You're triaging email for a founder. For each email, output JSON:
{ "category": "VIP" | "ACTION" | "FYI" | "NOISE",
  "priority": 1 | 2 | 3,
  "draft_reply": string | null }

Rules:
- VIP senders (hardcoded list, never override): see allowlist.
- ACTION means a human needs a response or a decision from me.
- FYI is informational. Newsletters, status updates, receipts.
- NOISE is cold outreach, automated re-confirmations, LinkedIn spam.
- Only write a draft_reply for ACTION. Match my voice from the examples.

Examples:

INPUT: "Hey, can you send over the Q3 deck before Thursday?"
OUTPUT: {"category":"ACTION","priority":1,
"draft_reply":"Sending now. Anything specific you want me to flag?"}

INPUT: "Your weekly Stripe summary is ready."
OUTPUT: {"category":"FYI","priority":3,"draft_reply":null}

INPUT: "I'd love to show you our new sales intelligence platform..."
OUTPUT: {"category":"NOISE","priority":3,"draft_reply":null}

Now classify:
[paste email]

That's the whole engine. The few-shot examples carry more weight than any rules paragraph. Add three more examples in your real voice and the drafts stop sounding like a chatbot.

The trust problem

The model will miscategorize something. It will mark a real customer as noise once. It will draft a reply that's almost right but slightly off. This is fine. What's not fine is letting it delete or send before you trust it.

For the first two weeks, AI does not auto-archive, auto-delete, or auto-send anything. It labels and drafts. You review. After you've watched it handle 100 plus messages and the error rate is honest, promote actions one at a time. Auto-archive noise first. Then auto-summarize FYI. Drafts stay as drafts for a long time, probably forever. The minute the model sends a reply you didn't read, you've lost the relationship layer that email is for.

Build trust slowly. The reason this works is the same reason it works in the CLI guide: you commit before letting the agent loose, and you read the diff after. Same loop, different surface.

The VIP guard

Whatever the model does, hardcode an allowlist of senders that bypass AI entirely. Your boss, your top five customers, your co-founder, your partner. These hit your real inbox with a notification, untouched, no classification, no draft. AI changes its mind every 200 tokens. A hardcoded list in a config file does not.

This single rule is what lets you sleep. The worst thing AI triage can do is hide a message that mattered. The allowlist makes that impossible for the senders who matter most.

Up next

You've got the inbox under control. The last guide in this pillar swaps email for the other recurring tax on your time: the weekly report nobody reads. Pillar 5 guide 3, "Replace your weekly report with AI," covers how to feed the model your week's raw activity and get back the update your team actually wants.