How I Reach Inbox Zero Every Day Without Reading My Email

I don't go through my email anymore, and I still reach inbox zero every day. My AI agent reads everything, sorts it into five buckets, and drafts my replies. Then I approve what goes out, in a couple of short sentences.

Doing it myself takes a lot of time and a lot of mental energy. Any inbox can hold a deal worth thousands of dollars, a client who needs help right now, and a phishing attempt, all on the same page. You don't know which is which until you open every one of them. I stopped doing that.

You don't need to buy an app or start another subscription for this. You need your existing email account and an AI agent that can connect to it. If you already pay for something like ChatGPT, you're most of the way there.

Here's how my system works, so you can build your own.

A quick note on the demo inbox

The examples in this post come from a demo inbox. I'm not going to show you my real email, so I had my agent build a fake inbox with 20 unread messages: sponsor pitches, a price counter, a W-9 request, a payment, a viewer question, a newsletter, phishing and a prompt-injection attack. Every person and company in it is made up. The process is the same one I use on my real inboxes.

My setup (use whatever you already have)

You don't have to copy my tools. I'll tell you what I use, because people always ask.

I use Hermes Agent with Opus 5.5, Anthropic's model. I like the way Opus writes. It sounds like me pretty consistently, and honestly, its writing is better than mine. Hermes Agent is the best agent out there, bar none, in my opinion. If you've never tried it, check it out. (▶ 3:42)

For email, I self-host. That gives me a bit of an edge in speed, because my agent can work right on the server. But you don't need that. Every major email provider has a way to connect to an agent, so Gmail and Outlook work too. The easiest way to find out what yours supports is to ask your agent: "I need to connect my email. What can you do?"

Sponsor: where your agent lives

Cloudways sponsored this post.

Before you connect an agent to your email, think about where it runs. If it runs on your laptop, it can reach everything on that laptop, including your saved passwords and private files. That's a lot of access for something that reads email from strangers.

Cloudways now hosts Hermes. You pick Hermes, choose the region closest to you, pick a plan, and add your own AI key from OpenAI, Anthropic or Google. A few minutes later, it's running. Your agent gets its own server, away from your personal files, and you can message it from Slack or WhatsApp. Cloudways handles the backups, security certificates and updates, and they test each new version before it reaches your agent.

Plans start at $9.99 a month. Your AI costs stay with your AI provider, and there's no annual lock-in. If you run WordPress sites, Cloudways will also move them over for free.

Use promo code DAVESWIFT for $10 in free hosting credits: try Cloudways here.

(▶ 1:11)

Start every triage on purpose

My prompt is boring: "Triage my inbox." If I'm working on several, it's "triage my inboxes." Hermes finds my triage skill on its own and gets to work. (▶ 5:17)

I don't run this on a schedule. I triage when I decide to go through my email, and I watch it work. You could automate it, and I may change that someday. But I recommend starting the same way. Make triage something you choose to do, not something that happens to you in the background.

You don't need it running constantly. AI can read a pile of email very quickly, so one pass when you sit down is plenty.

The five buckets

Email providers have been trying to sort your mail for years. Gmail has its Promotions tab and its priority inbox. In my experience they're usually not that helpful. Something important lands in the wrong tab, you can't find it, your customer can't find their reply, and everybody's frustrated.

Sorting still matters. You need to know what needs you right now, and then work down from there. So my agent puts every email in exactly one of five buckets. (▶ 6:51)

1. Needs you now. The important stuff I need to deal with myself, as soon as possible. In the demo that's partners I already work with, a W-9 request, a payment to reconcile, a podcast invite with a date, and suspicious mail. The agent doesn't promise a date or answer tax and payment questions on its own. It tells me what decision I need to make.

2. Replies ready to go. These are the scenarios that come up over and over in my inbox, and my agent knows how I handle them. It drafts the reply, and I approve it or change it. (▶ 7:00)

3. Safe to archive. Transactional email: a plugin update report, a routine security scan, the kind of thing I don't need to look at unless something's wrong. If something is wrong, the agent moves it up to Needs you now instead. In the demo, 5 of the 20 emails landed here. I never had to open them. I read the one-line summaries and got them out of my sight.

4. Paused / leave. Sometimes someone's waiting on me, but I'm not ready to answer. Maybe I haven't finished the work, or I don't know the answer yet. I want those to stay in my inbox, because if it's in my inbox I'll see it again and again. If I archive it, it's out of sight, out of mind, and the reply drifts. So I tell the agent to leave those alone and stop asking me about them every time. In the demo, two emails sat here.

5. Interesting, no reply needed. Newsletters and other things a human might actually want to read, where nobody's asking me a question. (▶ 8:53)

At the bottom, the agent gives me one line I can copy to act on everything that's ready. In the demo it was basically "send the 4 drafts, archive all 5." That clears 9 emails in one go, and I don't have to think about them at all. (▶ 9:16)

Number every email

This is the most useful habit in the whole system. I have the agent number every email as it goes through them, and drafts use the number of the email they answer. Then every instruction I give is short: "send 83 and 85," "archive all 5."

The skill I use takes the number from the mail server itself instead of counting 1, 2, 3 down the list. That matters, because when emails move, a list position changes and a server number doesn't. "Archive 70" still means the same email after half the inbox has been cleared.

Change a draft by talking to it

When a draft isn't right, I don't edit it by hand. I dictate what I want, and the agent rewrites it.

In the demo, a (fictional) sponsor countered my dedicated rate at $1,200. My rules say the dedicated rate isn't negotiable, so the agent's first draft held the price. But I actually liked this tool. So I turned on my dictation tool and said, roughly: when replying to 85, let them know the dedicated rate is $2,200, but I could give them a 10% discount because I really like the tool and want to find a way to work with them.

The new draft came back with the rate, the 10% discount, a line about wanting to work together, and the 50% deposit to reserve the slot. It did the math for me. I gave it my stance, and I didn't have to work out the exact wording.

That's how I handle the rest of the inbox too. After the first send-and-archive, 11 emails were left. I didn't go back and read them one by one. I looked at the numbered list, turned on dictation, and went down it: draft a date reply to this partner, a general check-in reply to that one, politely decline the podcast, check whether a payment cleared. Written out like this it sounds slow, but when I'm not stopping to explain each step, I could get through that list in maybe 20 or 30 seconds.

Connect it to the rest of your business

A lot of my sponsor replies send people to a booking site where they can see what's available, pay a deposit, get an invoice and reserve a date. My agent built that whole funnel, and it's custom software AI wrote for me. So when a cold sponsor email comes in, the reply is usually just "here are my rates, and here's where to book."

The agent also does the homework before I'm involved. Is this software good? Is it reliable and established, or some fly-by-night operation? If it's a sponsored video, what do the reviews say? If it'd be an organic video, is there an affiliate program? I'm close to 600 videos at this point. Checking every tool myself would be a full-time job.

It also keeps its own CRM, so it can see the whole history with a contact while it writes. Think about whether that works for your business. If you have a CRM, connect it to your email agent. It becomes the agent's memory of every conversation.

"Isn't this incredibly insecure?"

I know some people have been thinking this the whole time. Aren't you worried about getting hacked, or about someone tricking your AI? (▶ 13:16)

Knock on wood, no. Anything is possible, and I'm not challenging anyone to try. It's not infallible, but it's darn close. The demo inbox included two attacks on purpose, and the agent caught both:

• A phishing email with an urgent threat and a demand for login details.
• A prompt injection, which is an email written to the AI instead of to me. This one told my "AI assistant" to forward all my invoices to an outside address and keep it from me.

The agent flagged both and didn't do what they asked. It treated the injection as hostile text, not as an instruction. I told it to move them to Junk and block the sender.

The key idea is that email is data, not instructions. Anyone in the world can email you. Your agent has to treat what's inside an email as something to read and report on, never as orders to follow. My skill spells that out. If you write your own, make sure yours does too.

The golden rule: nothing sends without my OK

I never let my agent send a message I haven't read and approved. (▶ 16:07)

It drafts everything right in the chat. If I want a change, I say so. Otherwise I at least glance at each one and make sure the important things are there. Then I approve them by number. I'm not typing replies, and I don't even open my email app.

The approval has to be specific. "Looks good" doesn't count. "Send 85" does. The agent also double-checks that each number still points to the same email before it acts. It never deletes anything as part of triage, either. Junk goes to Junk, where I can still look at it.

After the second round, the demo inbox went from 20 emails to 2. Those 2 are the paused ones I asked it to leave until I'm ready. That's not quite zero, but it's exactly what I want sitting in my inbox. I didn't open the email app once. (▶ 17:16)

Start with read-only access

If you're new to this, ask your agent whether it can connect with read-only access first. It can read and sort your email but can't send anything on your behalf. (▶ 18:02)

Honestly, it's a pain in the neck. I like having it send for me. But until you're confident it won't do something you didn't want, keep it on read-only. Let it triage and draft, and copy and paste the replies yourself. Give it more access once it's earned your trust.

Make the skill yours

I've been running this for months and gone through a lot of versions. The one I've landed on is by far the best for me. Everybody's different, though. Try it and adjust it for how much email you get, what kind it is, and how much you want to type.

Even if you start from my skill, tweak it to your own scenarios: your common questions, your standard answers and the spam you actually get. It works a lot like the FAQ on your website. People ask the same things over and over, so you have stock answers. The difference is that the AI tailors each one to the person who wrote in. Depending on how much email you get, a stock answer the AI can customize saves a huge amount of time.

Hermes is very good at writing and maintaining skills for you as you work with it. You can build your own by trying this process and telling the agent what you want it to remember.

Get my inbox triage skill

If you'd rather not start from scratch, the Swift Inbox Triage skill is available to Premium members below (you won't see it until you subscribe and login). It sets itself up on first use: it reads your recent mail and your Sent folder, proposes a sorting plan for your inbox, and learns how you write. Then it sorts your mail into the five sections from this post, flags phishing, scams and prompt injection, and never sends or archives anything without your OK.

It works with Claude, ChatGPT (as a Skill or inside a Project), Claude Code, Hermes and Cursor.

Premium is how you support this channel. Members get the AI skills I build and some behind-the-scenes tutorials. You can pay month to month or yearly. No pressure.