Inbox triage with Claude: from 200 emails to 20 in 15 minutes

If you open Monday with a four-digit unread count, this is the cheapest piece of automation you can build. No API keys, no IT ticket, no new SaaS. Fifteen minutes of copy-pasting buys back an hour.

The problem with "AI for email"

Every email vendor is racing to bolt an AI assistant onto your inbox. They auto-summarise, auto-draft, auto-reply. They also auto-hallucinate, auto-leak, and auto-bury the one message that mattered. I tried six of them. I went back to copy-paste.

The pattern below keeps you in the loop where it matters (deciding what's important, hitting send) and offloads the part you actually hate (reading 200 subject lines and triaging them).

What you'll need

• An LLM you trust with internal content — Claude, ChatGPT, Gemini, or your enterprise equivalent
• A way to export the last 24–72 hours of email subjects + senders + first 200 chars (most clients let you paste the inbox view directly)
• 15 minutes

Step 1 — Dump the inbox

Open your inbox view and select the unread block. Copy it. Paste it into the model in this shape:

From: Priya at vendor.example   Subject: Q2 invoice attached
From: Manager at company.example Subject: Re: deck for Thursday
From: Stripe notifications       Subject: Receipt for $42
...

One line per email. That's it. No bodies needed for triage — sender + subject is enough signal 80% of the time.

Step 2 — The triage prompt

Paste this above the dump:

You are my inbox triage assistant. Cluster the emails below into these buckets, in order:

1. **Reply today** — anything where a person is waiting on me and the cost of delay is real (commitments, blockers, deadlines this week).
2. **Reply this week** — needs a response but not urgent.
3. **Read later** — newsletters, FYI threads, internal announcements.
4. **Trash** — receipts, automated notifications, no-action marketing.

For bucket 1, also draft a one-line action for each ("approve the budget", "ask Priya for the revised invoice", etc.). Do not draft full replies yet.

If you're unsure where something goes, put it in bucket 2 and flag it with a (?).

Why this works: you've moved the decision boundary from "which 200 messages?" to "did Claude get these 20 right?" That's a 10× cognitive saving.

Step 3 — Spot-check, then act

Scan bucket 1. If anything looks miscategorised, drag it. Then for each item in bucket 1, ask:

Draft a 3-sentence reply to the email from {sender} about {subject}.
Tone: warm but direct. I'm British-English. End with a clear next step.
If you don't have enough context to reply, ask me one clarifying question instead.

That last sentence is the one most people skip. Without it, the model will confidently invent details. With it, you get a clean "what does Priya mean by 'the revised version' — the one from last Tuesday or the one she sent this morning?" and you've avoided a wrong reply.

What this won't do

• Reply to anything sensitive. HR, legal, personal. Don't paste it in. Skim those yourself.
• Catch sentiment shifts. If a client is quietly fuming, the LLM will see "Re: project status — quick check-in" and route it to bucket 2. You still need to read.
• Replace your inbox rules. Filters and labels still beat this for predictable, repeating traffic.

What changed for me

Two weeks after I started doing this on Monday and Thursday mornings: Monday's panic-skim turned into a 20-minute coffee. Bucket 1 averages 14 items. I reply to all of them before standup. The other 180 emails get one batch operation — archive or read-later — and nobody has noticed.

The unsexy truth about AI productivity gains: they almost never come from a clever new tool. They come from finding the one task you do every week that's mostly low-stakes pattern matching, and outsourcing just that part.