The agent that reads my 200 emails
for me,
every morning.
I named it Hermes. It's a little assistant I built to open my inbox every morning at 7am, spot the three or four genuinely urgent messages, and file the rest away. Four months of use, 91% accuracy, around 4 hours saved a week. And yes, I'll also tell you about the day it replied to my mom like she was a sales prospect. Step-by-step tutorial for someone who has never written a line of code.
What you'll learn
- What Hermes does — it opens my inbox, spots the genuinely urgent messages, files the rest into three or four folders, and writes reply drafts for the easy cases. I read them over, I click, I send.
- The measured payoff — over four months of use, roughly 200 emails a day processed, 91% filed correctly, 4 hours saved a week on sorting. Not a marketing promise, my real numbers.
- How to build it — an account with the AI (Claude or ChatGPT), a connection to Gmail, a piece of text that explains your rules to the agent. Plan on a Saturday afternoon for the first version.
- Who it's useful for — an inbox under siege (50+ emails/day), wanting your mornings back, comfortable copy-pasting. If you get 10 emails a day, don't waste your time.
- My honest take — it's the tool that saved me the most time in 2026. But I had three serious mishaps along the way, and I'll tell you all of them so you don't repeat them.
- Transparency: article written with Claude (Anthropic), reviewed and approved by me. Hermes itself runs on Claude. If you spot a mistake, write to me, I'll fix it.
Four months with Hermes.
Before I sell you the magic, I want to tell you how it actually went. Not the "everything's great" version from the YouTube videos. The version where I struggle, where I break things, and where I eventually figure out what works.
December 2025. I'm averaging 200 emails a day — between the notifications I let pile up, the newsletters I never really read, the actual clients writing to me, the cold-email pitches (prospecting emails you get without ever asking for them). I spend about 1 hour each morning sorting. Times 5 days, that's 5 hours a week. Over a year, that's the equivalent of six full-time weeks of work. To sort emails. Help.
I decide to build a little assistant. I'd never done this before. I'm an entrepreneur, not a developer. But in 2026, you can build this kind of tool with an AI holding your hand (I used Claude, the AI from Anthropic, to help me write every line — I'll explain how further down). The first prototype takes one Saturday afternoon.
The real numbers after four months
The three instructions that actually worked
The agent doesn't do magic. It does exactly what you tell it, as long as you tell it well. Here are the three rules I ended up locking in after plenty of failed attempts. Keep them in mind, we'll come back to them in the tutorial.
Rule 1 — "Urgent" only if it's genuinely urgent
Label as "urgent" only if the sender asks me for a concrete action within 48 hours AND it's a paying client. The "AND" is crucial — without it, the agent was labeling 80% of emails as urgent. With it, we dropped to 5%, which matches the reality of my inbox.
Rule 2 — "Later" for newsletters I never open
Label as "read later" any newsletter I've opened fewer than 2 times out of the last 5 sends. That's honest sorting: if you haven't bothered to open it 3 out of 5 times, it isn't a priority. I unsubscribe in batches once a month.
Rule 3 — "Soft spam" for disguised cold pitches
Label as "soft spam" any email whose subject line has more than 2 emojis OR a discount percentage. Surprisingly accurate. These are the pitches that slip past the usual spam filters but have no business being in your main inbox.
And then the day Hermes replied to my mom
I have to tell you this one. After two months, I think it's doing such a good job sorting my email that I decide to give it a new power: writing and sending the simple replies on its own. Not just preparing a draft — actually sending. For cases where it's just "ok thanks", "got it", "let's talk again next week". You know the kind.
Day one. All good. Day two, I see an email from my mom go by in the history. She writes to me: "Hi sweetheart, are we doing lunch on Sunday with your brother? Love, Mom." Hermes's reply: "Dear Ms. Sagnier, thank you for your proposal. Could you let me know your more precise availability so we can lock in a slot? Best regards, Jérémy."
The full tone of a salesperson talking to a hot lead. To his own mom. I shut off the auto-send feature within the minute. I called my mom to laugh about it, but she didn't laugh all that much. And I added a new rule: the agent prepares drafts, it never sends on its own. Ever. Unless I click. That's my scar. Learn from mine, not from yours.
What you take away from this story
An agent does exactly what you tell it, with no common sense. If you give it the power to send, it'll send. If you give it the power to delete, it'll delete. Until you're 100% sure, keep it in "suggest, I click" mode. Active mode only on truly reversible actions, like labeling or archiving.
Three useful things it does for me.
When I say "agent that sorts my email", that stays vague. Here are the three concrete cases Hermes solves for me every day. If one of them resonates with your day-to-day, it's probably worth your Saturday afternoon.
The morning sort
Separating the genuinely urgent from the noise
Before: I open Gmail, I see 47 unread emails, I skim them all, I decide on the fly, sometimes I miss an important email buried in the noise. After: Hermes has already labeled everything by 7am. I open the "urgent" view, I see 3 or 4 emails, I handle them in 10 minutes. The rest waits for the afternoon or the next day.
Measured payoff: 45 minutes every morning. That's the equivalent of a free lie-in every week. Over four months, that's exactly 60 hours of my life I didn't spend sorting email. That's a long weekend won back.
Setup cost: a Saturday afternoonThe easy drafts
Writing the "ok thanks, got it" replies
For simple confirmation emails (a client saying thanks, a supplier sending an invoice, a partner acknowledging receipt), Hermes prepares a draft in Gmail. I open it, I read it over, I send. Three seconds instead of thirty. For 20 emails a day of this kind, that's 9 minutes saved every single day.
The key point: it prepares, it never sends. See the incident with my mom above. Always confirm with a click, never auto-send. This rule is non-negotiable.
Volume handled: ~ 20 drafts / dayThe big tidy-up
Filing incoming emails into themed folders
I have three big active folders: "clients", "suppliers", "watch". Hermes files each incoming email into the right one. When I'm looking for something three months later ("what was the designer's invoice from March?"), I go straight to "suppliers" instead of digging through the whole inbox.
The hidden benefit: my inbox now only holds recent, untreated stuff. No more of that feeling of never seeing the bottom of the ocean. It's mental, it's invisible in the numbers, but it's probably the most valuable part.
Average accuracy: 91%How it works, really.
You'll see, it isn't magic. If you understand how a human assistant works, you'll understand Hermes. It's exactly the same logic, just with a computer in place of the brain.
The analogy that clicks
Imagine you hire a new assistant. On day one, you tell them: "Here's my inbox. Here are my rules: an email from this client is urgent. A newsletter goes in 'later'. A cold pitch, you archive. " You take 30 minutes to explain it, you show them ten concrete examples, you tell them what you expect.
The next day, your assistant shows up at 7am, opens your inbox, applies your rules. When they're unsure, they set the email aside for you to decide. After a week, you correct two or three things ("no, that client isn't actually urgent"), and it's dialed in.
Hermes is exactly that. Except your "assistant" is an AI. And you don't talk to it, you write its rules in a piece of text we call "the system prompt" (the equivalent of the assistant's employment contract — the list of rules it must follow all the time, in every case).
The three ingredients (and that's it)
What Hermes needs to run
- A brain — the AI itself. You pick one (Claude, ChatGPT, or a free AI installed on your own machine). It's the one that reads each email and decides.
- Tools — the connection to Gmail. That's what lets the AI read emails, add labels, prepare drafts. Without it, the AI is blind.
- Clear instructions — the system prompt. The 20-to-50-line text that explains your rules to the AI. That's where everything is won or lost.
When you start Hermes up, here's what actually happens behind the scenes:
- The agent opens your inbox (via the Gmail connection) and grabs the unread emails.
- For each email, it sends it to the AI saying: "Here's this email, here are my rules, tell me which label to stick on it."
- The AI replies: "Label = urgent" (or read later, or watch, etc.).
- The agent applies the label in Gmail (via the same connection).
- If it's an email that deserves a reply draft, the AI writes it, the agent drops it into Gmail as a draft. You see the little "draft" tag appear in Gmail in the morning.
That's it. No magic. No advanced robotics. Just an AI that reads text, applies your rules, and ticks boxes in Gmail for you.
Why I named it Hermes
Hermes, in mythology, is the messenger of the gods. The one who carries the news between worlds. It was an obvious fit for an agent that spends its life in my inbox. You can rename it whatever you want — Jeeves, Alfred, Stuart, Postman Pat — it changes nothing about what it does. It's just nicer to say "Hermes did its sort" than "the Python script ran overnight".
Let's build your Hermes.
Here's the path to follow, exactly how I'd redo it today in 2026. Plan on a Saturday afternoon for the first version, plus an hour of tweaking the following week to fine-tune the rules. I'll use Mac as the running example because that's what I have. On Windows or Linux, it's exactly the same logic, I'll flag it whenever it changes.
Before you start — the honest version
This tutorial asks you to copy-paste, create 2 accounts, and fill in a text file. If you've never done that, it's doable. If you're not at all comfortable with a computer, you're better off reading the "Without installing anything" block right at the bottom — you'll get 70% of the result in 2 minutes, and you'll come back here when you're ready to push further. No shame in that.
Create an account at Anthropic (the maker of Claude)
The AI that's going to do the work is Claude. It's the one that gives me the best results on this kind of sorting (I also tried ChatGPT, it's great too, pick whichever you already know). Go to console.anthropic.com. Click "Sign up", enter your email, verify via the link you receive.
Once logged in, you go to "API Keys" in the left-hand menu. Click "Create Key". Give it a name (for example "Hermes mail"). You get a long key that starts with sk-ant-.... Copy it and keep it somewhere safe — you'll never be able to see it again afterward. Paste it into a local note on your computer, never in an email.
Choose how you connect Gmail
You have two routes, to pick based on your comfort level. Here's the honest comparison, no spin.
Route A — Make or Zapier (the "no-code" option). You create an account on make.com (or zapier.com). These tools are for wiring services together by drag-and-drop, without writing a single line of code. You connect Gmail inside (a click, a Google authorization), then you connect Anthropic (by pasting your key). Plan on 30 minutes to get the hang of the interface. Free up to 1,000 actions a month — more than enough to get started.
Route B — A little Python script (the "I like to understand" option). You install Python on your Mac (it's already there, actually), you copy-paste a 50-line script I give you below, you run it every hour with a system command (cron on Mac/Linux, Task Scheduler on Windows — the equivalent of an alarm clock telling your computer: "every hour, run this script"). More control, but it means diving into a terminal.
My advice for your first Hermes: Route A. You'll have something that works in 2 hours. You can always migrate to Route B in 6 months if you want more finesse.Connect Gmail (5 minutes, Route A as the example)
In Make, you create a new "scenario". The first block is "Gmail · Watch Emails". You click, you authorize Make to connect to your Gmail (Google shows you a permission screen, you accept). You pick "Inbox" as the folder to watch. You set "every 15 minutes" as the frequency.
The second block is "Anthropic · Send Message". You paste your API key, you pick the claude-haiku-4 model (the fastest and cheapest for this kind of task). The third block is "Gmail · Add Label", which takes Claude's reply and applies the matching label.
Write the instructions (the system prompt) — the part that changes everything
This is where you teach your agent who you are, what you expect, and how you work. Your 91% sorting accuracy is 80% down to this text. Take 30 minutes to write it carefully. Here's the template I use, to adapt to your case.
You are Hermes, my assistant who sorts my inbox.
WHO I AM: Jérémy Sagnier, an entrepreneur based in Nice.
My real clients are: [list 3-5 names or email domains].
My close family uses these emails: [list].
My usual suppliers: [list 3-5 domains].
FOR EACH EMAIL, ASSIGN A SINGLE LABEL:
- "urgent": action requested within 48h AND the sender is
a paying client. Otherwise no.
- "client": email from one of my clients listed above,
not urgent.
- "family": email from a family member listed above.
- "supplier": invoice, quote, or exchange with
a listed supplier.
- "watch": newsletter or useful AI / business news.
- "later": newsletter I rarely read, non-priority
information.
- "soft spam": commercial cold pitch, more than 2 emojis
in the subject, or a discount percentage.
IMPORTANT RULES:
1. If you're torn between two labels, pick the lower-priority
one ("later" rather than "urgent").
2. You NEVER reply to emails automatically. You
prepare drafts only for: simple confirmations
("ok thanks"), acknowledgments, polite declines
of cold pitches.
3. If the email is from my mom, my sister, or my brother,
you write NO draft. I always reply to them
by hand. (Little inside joke: you know why.)
REPLY FORMAT: return just the label word,
in lowercase. Nothing else. No explanation sentence.
You replace the [list...] placeholders with your real info. You copy all of this text into Make's Anthropic block (the "system prompt" field). The email to analyze arrives as a "user message" block via the Gmail variable.
The secret is precision. "Urgent if paying client AND request within 48h" gives 91% accuracy. "Urgent based on your gut" gives 80% of emails labeled urgent. The simple rule: no vague adjectives ("important", "relevant", "for you"), measurable criteria instead.Test on 10 emails in "watch" mode
Before letting Hermes label your inbox for real, you're going to run it in test mode. In Make, you click "Run once" ten times in a row, on your 10 most recent emails. For each email, you look at: which label does Hermes suggest? You compare it with what you'd have picked yourself.
If Hermes gets it wrong on 1 or 2 emails out of 10, that's normal and acceptable, you'll refine the rules next week. If Hermes gets it wrong on 5 out of 10, don't turn on active mode — go back to step 4, your system prompt isn't precise enough. Add concrete examples: "example: an email from support@stripe.com saying 'your payment has arrived' → label = supplier".
Keeping an error log helps. For 2 weeks I logged every mistake Hermes made in an Apple note. After 14 days, I spotted the patterns and added 3 precise rules to the prompt. Accuracy went from 78% to 91% in one afternoon of tweaking.Turn it on for real (labeling-only mode)
When your 10-email test looks good, you turn the scenario on in Make (big green "ON" button at the top). From now on, every 15 minutes, Make reads your new emails, asks Claude to label them, and applies the label in Gmail. You have nothing left to do.
Keep it in labeling-only mode for at least two weeks. Meaning: it labels only, it archives nothing, it deletes nothing, it sends nothing. You confirm each day that the labels it applies suit you. If they do, you can move on to the more active features afterward.
Watch your Anthropic credits the first week. You'll see exactly how much it consumes. For 200 emails a day with Haiku 4, I spend about €0.25 a day. No nasty surprises.Add reply drafts (only after 2 stable weeks)
If after 2 weeks you're happy with the sorting, you can move up a level: let Hermes prepare drafts for simple emails. You add a new block to your Make scenario: "Gmail · Create Draft". The draft shows up in Gmail with a little "Draft" tag — you open it, you read it over, you send (or you bin it).
You tweak your system prompt to add an instruction: "For emails labeled 'supplier' with a simple confirmation, write a short draft: 'Hi, got it, thanks. Have a good day. Jérémy.'". Hermes prepares the text, Make drops it into Gmail as a draft. No auto-send. Ever.
Remember the incident with my mom. Draft mode = OK. Auto-send mode = NEVER. The rule holds even when you push your agent to 6 months of use and trust it completely.Keep an eye on it during the first month
Block 5 minutes every Friday to open Make's history and look: how many emails were processed this week? How many visible mistakes? Is there a type of email that systematically lands in the wrong category?
After a month, you'll have sorting that holds at 90%+ accuracy, and you'll have your mornings back. After that, the tool runs quietly for months, with nothing to touch. Maintenance is 5 minutes a month at the very most.
On the evenings when Hermes goes down (Make is having an outage, Anthropic's API is overloaded), you get an email from Make warning you. It's rare — over 4 months I had 2 incidents, both fixed in an hour without me lifting a finger.And there you go, you've got your Hermes
If all went well, you now have an agent that opens your inbox every 15 minutes, labels automatically, prepares a few drafts, and gives you 45 extra minutes in the morning. You can do the same for your calendar ("book a slot for this request"), your client contracts ("read and summarize them for me"), your invoices ("file them by month"). But start with a single case. Master it. Then you add more.
The four mistakes I made.
I'll list them in the order I lived through them. None is catastrophic, but each cost me time you can avoid. The general rule: start small, watch a lot, only turn on a new feature after two stable weeks on the previous one.
Pitfall 1 — Sending your email history to an AI without thinking
You might be tempted to send your last 5,000 emails to the AI so it can "learn your style". Stop. When you send an email to Claude or ChatGPT, that data goes through their servers. Your contracts, your client info, your sensitive exchanges — all of it. For Hermes, I only send one email at a time, just the subject line and a small snippet of the body, only for the sorting. Never the full history. If you want to do better, run locally with Ollama (I talk about it in the article on running AI on your own machine) — nothing leaves your computer.
Pitfall 2 — Wanting 10 email categories from the start
At first, you'll want to sort finely: "urgent client A", "urgent client B", "easy draft", "medium draft", "forward to Sophie", "invoices over €500", and so on. That's a guaranteed disaster. The AI hesitates, gets it wrong, your accuracy drops to 60%. Start with 3 categories max (for example: urgent / later / archive). Once you hit 95% accuracy on 3, add a 4th. Then a 5th. My current Hermes has 7 labels, but it took me 3 months to get there, in stages.
Pitfall 3 — Turning on active mode without testing on real emails
The temptation is strong: you write your system prompt, it works on 3 examples you made up, you click "ON" in Make. The next day, your agent has archived an important email from a client, you never see it, you lose the client. Three weeks later, you take everything apart to figure out why. Never skip step 5 (test on 10 real emails) of the tutorial. And keep labeling-only mode on for at least two weeks. It's the only rule that protects you from silent disasters.
Pitfall 4 — Forgetting Mom (and every human case)
My famous incident. The rule: never auto-send, even for cases that seem trivial. You can prepare drafts, that's great. But the human must always be the last link. An AI has no common sense — it can't tell the difference between "my mom inviting me to lunch" and "a partner asking for a meeting". List explicitly in your system prompt the cases where Hermes must never reply on its own (family, close friends, sensitive topics). And even then, keep draft mode. No auto-send. Ever.
Who it's really useful for, who it isn't.
I don't like articles that sell you a tool by saying "everyone should try it". Here's my decision matrix, based on 4 months of use and feedback from a dozen friends who built their own Hermes.
| Your profile | Verdict | Why |
|---|---|---|
| You get 50+ emails / day | Go for it | The return on investment is immediate. You'll save 1 to 4 hours a week depending on your volume. |
| You get 10-50 emails / day | If it's fun for you | The time saved is marginal (15-30 min/week). But it's a great first project for understanding how AI agents work. Build it to learn, not to save time. |
| You get fewer than 10 emails / day | Don't bother | You'll spend more time building it than it'll ever save you. It isn't for you, it's for inboxes under siege. |
| You work with ultra-sensitive data (lawyer, medical, professional secrecy) | Local route required | Not the Claude API. You install Ollama and a free model on your own machine, as explained in the AI-at-home article. Hermes exists in a 100%-local version, it's just a bit more technical to set up. |
| You're really not comfortable with a computer | No-code route | Go read the "Without installing anything" block at the bottom of this article. You won't get full automation, but you'll get 70% of the result with a copy-paste into Claude.ai. |
If you really don't want to code: three ready-made alternatives
If you read this article and think "okay, this is too much for me, I just want my email sorted", there are off-the-shelf tools that do part of the job. Honestly, they're less powerful than a homemade Hermes, but they work out of the box.
SaneBox
The classic. Automatic sorting of unimportant emails into a "SaneLater" folder you read whenever you want. Accurate, simple, but not very customizable. A good starting point.
Superhuman
The premium option. Ultra-fast interface, AI features to write drafts, smart sorting. Pricey, but loved by people who spend their lives in their inbox.
Spark Mail
Automatic inbox sorting into 3 sections: Personal / Notifications / Newsletters. No personalized AI, but a default sort that's often good enough. And it's free.
Microsoft Copilot in Outlook
If you're on Outlook (and therefore Microsoft 365), the built-in Copilot summarizes long threads, suggests replies, and sorts. Less sharp than a homemade Hermes, but zero installation.
My blunt take: if you just want to "stop being at your inbox's mercy", start with SaneBox (€7/month, 5 minutes to turn on). If you genuinely want to learn how to build an agent and the idea of a custom tool appeals to you, build your Hermes — it's 100 times more instructive and 50 times cheaper in the long run.
The 2-minute version.
If everything above feels like too much, here's the express version that gets you 70% of the result with nothing to install, no API account to create, no touching Make. It's what I recommend for the very first try.
Sort 10 emails in Claude.ai
1. Go to claude.ai. Create a free account if you don't have one.
2. Create a new "project" (button in the left-hand menu). Call it "Email sort".
3. In the project, put these instructions in the "Custom instructions" field:
You are Hermes, my assistant who sorts my inbox. For each email I show you, classify it as: - urgent (action within 48h, paying client) - later (newsletter, non-urgent info) - archive (cold pitch, already handled) Reply with a single line per email: "Email 1: urgent, because [short reason]"
4. Open Gmail. Copy the contents of your 10 most recent emails into the Claude chat (subject + first lines). Send.
5. Claude sends you back a sorted list. You do the sorting by hand in Gmail following its recommendations.
And what's next?
Once you've mastered your first Hermes, you've unlocked a transferable skill. You can run it again for all sorts of repetitive tasks in your life. Here are my ideas for what's next, in order of difficulty.
- Hermes-calendar — an agent that reads meeting requests in your email, cross-checks with your Google Calendar, and proposes 3 slots in the draft reply. 2 more hours a week saved for me.
- Hermes-invoices — an agent that reads the attachments in "supplier" emails, extracts the amount, the date, the name, and files it all into a Google Sheet. No more Sunday-night bookkeeping.
- Hermes-watch — an agent that reads 15 business/AI newsletters every morning, pulls out the 3 real bits of news, and sends them to you in a single digest email. That's exactly what I do for AI Playbook.
- Hermes-local — the same thing but without sending a single email to Anthropic. You install Ollama (free) and a model that runs on your own machine, and you plug Make into that local AI via a technical connection. For pros who can't send their data elsewhere. That's my next step.
If you haven't yet read my article that lays the foundations of running AI on your own machine (Ollama, models, how it works locally), it's right here: Running a real AI on your own machine. It's the natural companion to this article if you want a 100%-private Hermes. And if the idea of agents in general interests you, my article Claude Code loops explains the deep principle behind all the mechanics.
To go further on agents: The complete guide to AI agents sets out the theory, Build a Gmail agent reuses the same mechanics on another use case, and The agent that reads my contracts shows how to apply Hermes to work PDFs. You can read them in any order.
Three things to do this weekend
- Test the 2-minute version in Claude.ai on your 10 most recent emails. Without installing anything. You'll know in 5 minutes whether the idea speaks to you.
- If yes, build your first Hermes with Make and Route A of the tutorial. Plan on a Saturday afternoon for the first working version.
- Keep it in labeling-only mode for two weeks before touching anything else. Watch it, log the mistakes, adjust the rules. That's what takes accuracy from 78% to 91%.
And if you're stuck on a step, or you think I've got something wrong somewhere, write to me. I read everything, I reply. That's exactly the kind of feedback that made me improve my own Hermes over the months — and it's probably what'll get me to ship you a new tutorial in 6 months.
Hermes FAQ.
Exactly how much does an agent like Hermes cost?
About €8 a month to process 6,000 emails (200 a day) with the claude-haiku-4 model. That's €0.25 a day. Make is free up to 1,000 actions a month, more than enough to get started. No other hidden cost: no subscription, no license to buy.
Do my emails go through Anthropic when Hermes sorts them?
Yes — Hermes sends Claude the subject line and a snippet of the body of each email to classify. Anthropic has a default no-retention commitment on the API, but your data does pass through their servers. For ultra-sensitive data (lawyer, medical, professional secrecy), you need to switch to the local route with Ollama and a model that runs 100% on your own machine (see the AI-at-home article).
Does Hermes work with Outlook or only with Gmail?
Make and Zapier both have official Outlook and Microsoft 365 connectors. The tutorial follows exactly the same logic: you swap the Gmail block for the Outlook block, the rest is identical. If you're already on Microsoft 365, Microsoft Copilot for Outlook (~ €22/month) covers part of the need with nothing to install.
How long does the full tutorial setup take?
Plan on a Saturday afternoon for the first working version (creating the Anthropic and Make accounts, connecting Gmail, writing the system prompt, testing on 10 emails). Plus an hour of tweaking the following week to fine-tune the rules. In total: 4 to 6 hours to go from zero to a working Hermes.
What happens if the Anthropic API goes down in the middle of the night?
Make emails you the moment a scenario fails. Your emails stay in your inbox, simply unsorted until the service comes back. Over 4 months of use, I had 2 incidents (Make down or Anthropic API overloaded), both fixed in under an hour without me lifting a finger. No email lost.
How secure is Make for my email data?
Make stores processed data for 30 days by default for debugging. You can lower that to 7 days in the scenario settings, or turn on Data structures mode to store only the variables and not the body of the emails. Make is ISO 27001 certified and GDPR-compliant, hosted in the EU. For anything sensitive, the local route is a must.
How many tweaks does the system prompt need before it works well?
Plan on 5 to 10 iterations over the first two weeks. For 14 days I logged every mistake Hermes made in an Apple note, I spotted the patterns, and I added 3 precise rules to the prompt. Accuracy went from 78% to 91% in one afternoon of tweaking. After that, 1 change a month at most.
Can Hermes handle 1,000 emails a day without crashing?
Yes, easily. Make absorbs it with room to spare (10,000+ actions/month on a paid plan starting at €9/month), claude-haiku-4 processes each email in under a second. You'll pay around €40 a month on the Anthropic side for 30,000 emails. The limiting factor isn't technical but the accuracy of the sorting once the volume gets too varied.
What's the difference between Hermes and SaneBox or Superhuman?
SaneBox (~ €7/month) automatically sorts unimportant emails into a folder, with no deep customization. Superhuman (~ €30/month) offers a premium interface with pre-wired AI features. Hermes is custom-built, written with your own real rules, plugged into Gmail/Outlook, and costs €8/month. Sharper but it takes a Saturday to set up.
Do you need to know how to code to build your own Hermes?
No. Route A of the tutorial (Make + Anthropic + Gmail) doesn't write a single line of code: it's drag-and-drop in a visual interface. If you can copy-paste a piece of text and create 2 online accounts, you can do it. Route B (a Python script) is for the curious who want to understand how it works — and it's still just 50 lines of code ready to paste.
Shall we keep going?
I test AI for real and share what works, with no jargon and no hype. If this article was useful to you, the simplest way to never miss anything is my Friday letter. And if you have a question or a doubt: reply to me, I read everything.