In partnership with

AInauten logo

Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

A quick status report: the world isn't ending yet. Even if parts of last week made it look that way.

The stock market wiped out $2.5 trillion in tech value in one day. Crypto is staying afloat on HODL. Outlook was offline for almost a day, giving millions of people Inbox Zero. And on Friday, the skies over the US went quiet: no planes were allowed to take off.

Houston, do we have a problem? Nope. We have a newsletter. πŸ˜‰

And it's packed with prompts you can copy right away: one sorts your inbox by your rules; the other keeps your shopping agent away from fake stores on Black Friday. Plus, Anthropic is heading to the stock market while Claude keeps working its way through the office...

Here's what's inside:

Let's go!

Attio is the #1 fastest-growing CRM on Brex and Ramp

Built for the most ambitious companies in GTM, Attio is an end to end revenue platform that allows teams to run multiple plays across different revenue motions and lifecycles.

Teams at the frontier are already building on it: Granola, turbopuffer, Modal, and many many more.

πŸ“¬ How-to: Get your inbox to handle your email automatically

As we said, Outlook was offline for almost an entire day last week. For some people, it was the most relaxing workday of the year πŸ˜†. But the next morning, the inbox was twice as full...

Time to let an agent tidy up. By your rules.

One page of rules instead of Inbox Zero

The trick comes from a workflow that's making the rounds. It's so simple we're wondering why we didn't think of it ourselves.

Visual note: the animation is in German. The five steps are: check access, read your emails, propose rules, sort each morning, and let you decide.

Inbox assistant workflow: connect, read, propose rules, sort each morning, you decide

Start with a one-page operating manual for your inbox. Who's a VIP? What can go straight to the archive unread? What needs a summary? What needs your immediate attention? How do you handle meeting requests? And what do you sound like when you reply?

You don't even have to write the page yourself. The agent drafts it from your own correspondence.

Then the agent goes through your inbox each morning. But not inside the inbox itself!

It puts everything in a running document. One line per email: what it's about, what it suggests, and a draft reply where needed.

You respond there with one word or phrase. "Archive." "Decline." "Send meeting invite." "Follow up." The agent acts on it, with replies saved as drafts in your email system. Nothing goes out until you approve it or send it yourself.

The real work in an inbox isn't reading. It's all those small decisions. This workflow brings them together into a few minutes each morning.

Here's what the result looks like.

Visual note: this German screenshot shows the assistant drafting inbox rules, including VIPs, priorities, and archiving only with approval. Names are blurred.

The inbox assistant proposes an operating manual; names are blurred

The prompt to copy

No special tool required. This works with ChatGPT, Claude, and pretty much any email app you can connect through connectors or browser use.

What matters is what you allow the agent to do. Our starting recommendation: reading and creating drafts, yes; sending and deleting, no. You choose whether drafts go directly into your email system or into a document.

Copy this prompt into your tool of choice. The agent checks access, looks at how you email, and proposes an operating manual. If you want, it can then run automatically every morning. Over to you. Have fun!

Set up an inbox assistant for me. Only ask questions at the points below or if you get stuck.

1. Access: Check whether you can access my inbox (Gmail, Outlook, or another service). If not, tell me where to enable the connection. You need read and draft access, plus modify access for archiving.

2. Get to know me: Read my last 100 sent emails and the last 14 days of incoming mail. Work out whom I reply to quickly, what I ignore, what arrives automatically, how I schedule meetings, and how I sound.

3. Proposal: Turn that into a one-page operating manual: VIPs, urgent items for me, archive, summarize, meetings, and my reply style. Ask at most three follow-up questions.

4. Storage: Ask where the manual and a running inbox document should live (Google Doc, Word, Notion, or a local Markdown file), and whether reply drafts should go directly into my email drafts or into the document.

5. First pass: Add it at the top of the inbox document, dated: urgent items first, then one line per email with your proposed action and a link to the email. Save replies in my style where I specified in step 4. I respond to each line with a word or phrase ("Archive," "Decline," "Send meeting invite"), in the document or chat, and you carry it out.

6. Routine: If your tool supports recurring tasks, propose a daily pass at 8 AM in my time zone and set it up after I say yes, without duplicates. Confirm the time and next run. From then on, process only new emails. If this isn't possible, clearly tell me that nothing is running automatically.

7. Learning: If I repeatedly give the same response, propose it as a new rule. Never change rules silently.

Nothing is sent, deleted, or archived without my approval, including during automatic runs.

Free, private email that puts your privacy first

Proton Mail’s free plan gives you private, end-to-end encrypted email with no ads and no data selling. Built by scientists and privacy advocates, it keeps you in control.

This Black Friday, get 60% off paid plans for extra storage and premium features.

πŸ›’ Your agent goes shopping. Can it spot the fake store?

Black Friday is less than two weeks away.

This year, many people aren't asking Google anymore. They're asking ChatGPT and others: "Where can I get the cheapest Birkenstocks?"

A new study asked ChatGPT, Gemini, and Perplexity nearly 6,800 shopping questions about 151 brands. Risky stores appeared in all three: counterfeit shops imitating real retailers, old brand domains that now belong to someone else, and dead sites just waiting for scammers to buy them.

Visual note: parts of this animation are in German. Risky stores appeared in 15% of Perplexity answers, 8% of ChatGPT answers, and 4% of Gemini answers; warnings appeared in only 9% of cases.

AI shopping study: risky stores in 15% of Perplexity, 8% of ChatGPT, and 4% of Gemini answers; warnings in 9% of cases

ChatGPT was the most likely to warn users, on almost one in four risky links. Across all three bots, warnings appeared in just 9% of cases. Questions about country-specific stores were the most vulnerable. Exactly the question we often ask in German-speaking Europe: "Is there a German store for...?"

Sometimes the chatbots even invented regional stores for brands that don't have any. And small brands were affected twice as often as big ones.

To be fair, the study comes from a company that protects brands against counterfeiting. No reason to panic, then. But caution still makes sense.

Now the agent can pay, too

Meanwhile, agents are getting wallets.

OpenAI launched ChatGPT Wallet on Friday. Your Dot, your personal agent in ChatGPT, uses it to shop for you. It never sees your real card number, and every purchase requires Face ID or a fingerprint. Later, a budget and category are supposed to be enough for the agent to order on its own. For now, it's US-only, on the Pro plan.

Visual note: the German labels describe the Wallet flow: state what you want, Dot suggests a gift and prepares the order, then you confirm with Face ID. Your real card number is not shared.

ChatGPT Wallet: the agent suggests a gift and you confirm with Face ID

Claude can't pay yet, but it can analyze your spending. Grok Bot goes further: it already shops online, using a one-time card number through Stripe, with your approval for each payment. Both are US-only too. Of course...

Stores are opening the door as well. Shopify, the platform behind millions of online stores, has brought its merchants directly into ChatGPT, Copilot, and Google's AI Mode and opened its catalog to more agents. Since late September, agents can even fill out and submit checkout themselves, once you've confirmed.

Convenient for you. Unfortunately, for fake stores too: an agent can't automatically tell whether a store is legitimate. If it picks the wrong one, even the best biometrics won't help. You'll just confirm the wrong purchase very securely.

Prompt/Skill: The fake-store check for your agent

For your Black Friday shopping, that means: let AI search, then have it check the store before you pay. Every time, not just when you happen to remember this newsletter.

Visual note: this German graphic summarizes the five checks detailed in the English Skill below: domain, company details, scam lists, external reviews, and payment.

Five fake-store checks: domain, company details, scam lists, reviews, payment

That's why this check isn't a prompt you'll forget next week. It's a Skill. Set it up once, and your agent uses it automatically whenever buying comes up. If your tool doesn't support Skills, put the text in its custom instructions or a project.

Create this as a Skill.

---
name: online-shop-check
description: Use automatically before every purchase and store recommendation. Checks whether an online store is legitimate before an order or payment.
---

# Online Store Check

Before I buy anything, you recommend a store, or you place an order: check every store. Only then proceed to the cart or checkout.

## Checks
1. Domain: The official brand website or an authorized retailer? Warning signs include additions such as "official," "outlet," or "-de," spelling mistakes, and newly registered domains.
2. Fake-store lists: The German consumer advice center's Fakeshop-Finder (verbraucherzentrale.de/fakeshopfinder-71560) and Austria's Watchlist Internet (watchlist-internet.at/liste-betruegerischer-shops). These resources are in German.
3. Reviews outside the store: Trustpilot, forums, Reddit. Many poor reviews or suspiciously uniform ones are warning signs.
4. Company details: A real business with an address, company registration, and phone number?
5. Payment and price: Upfront payment only, overseas bank transfers, crypto, or discounts above 50% are warning signs.

## Result
For each store, give a traffic-light rating (green, yellow, red), with one sentence per check and the link you checked. Explicitly say what you couldn't verify.

## Rules
- Yellow or red: don't order, pay, or enter any data. Ask me first.
- Red: suggest the official store or a well-known retailer.
- Never pay without my explicit confirmation.

With the Skill running in the background, you can even give your agent the whole wish list. It checks every link and builds a table with traffic-light ratings. It takes a few minutes and could save you real trouble.

Language note: the two scam-checking resources below are in German.

And if you'd rather check yourself: the German consumer advice center's Fakeshop-Finder is free, and Watchlist Internet lists almost 18,000 known fake stores.

Our take: The agent searches, you pay

Comparing options, summarizing specs, finding alternatives: AI already does that really well for us. But the final click stays with us for now, especially for small brands and country-specific stores, where error rates are highest.

We don't rely on the agent alone, either. We also run Bitdefender Web Sage in the browser. The extension uses AI to flag suspicious stores, fake ads, and fraudulent banking pages.

We show the rest of our security setup in the Black October Deep Dive. Usually Premium, it's currently freely accessible to you.

🧾 Claude builds its own Office, Anthropic goes public

Last week, Anthropic effectively introduced its own Office. Except you don't open programs inside it. You talk to Claude.

  • Excel is now Claude Sheets.

  • Miro is Claude Whiteboards.

  • The dashboard is Claude Data, with live data from your tools.

  • And instead of taking minutes, Claude Meetings joins the call and builds the documents while you're still talking.

Claude can read and edit all these files, and you can open them yourself at any time. Sheets, Whiteboards, and Data are available on all paid plans; Meetings is currently Team and Enterprise only.

Visual note: the German labels show Claude building a whiteboard, spreadsheet, and live-data view. The underlying demo is from Anthropic.

Claude Whiteboards, Sheets, and Data in action; excerpt from the official Anthropic video

Try it: Plan the holiday party in three prompts

Claude's list of new features is long. Just before an IPO, you like showing what you can do. Try it on a real project, such as a holiday party for your team, club, or family. One chat, three steps:

  1. Plan: Dictate whatever comes to mind about the party. Date, location, who's coming, who's bringing what, what's unresolved. Then: "Turn this into a board with a timeline, owners, and open questions." We use Claude Whiteboards for that.

  2. Budget: "Build a spreadsheet with all expenses, a per-person budget, and a traffic-light warning if we're over budget." Share the spreadsheet so everyone can enter what they're bringing and what it cost. That's Claude Sheets.

  3. Invitation: Open Canva beside the chat. "Use a holiday template to create the invitation, with the date and location from the board." You see every change live and approve it. This works through interactive connectors.

The clever part isn't any single feature. What you say in step 1 appears on the invitation in step 3, without copying anything. The workflow is similar in ChatGPT.

Honestly, Sheets and Whiteboards are still in beta, paid-plan only, and deliberately simple, without pivot tables or macros.

ChatGPT and Claude: From app-hopping to a cockpit

ChatGPT has very similar capabilities. Third-party apps have been running inside chat as Plugin Extensions for a few weeks. We're already using ChatGPT Meetings regularly as an alternative to tldv and Granola.

The chat is eating the apps, we wrote in October. And OpenAI's superapp wants to be your operating system. Now Claude and ChatGPT are turning into cockpits from which you control everything. Awkward for Microsoft: Copilot puts AI into Office; Anthropic puts Office into AI.

In three ways:

  • Connect: Your apps and data connect directly through connectors, MCP, or APIs. The chat retrieves what it needs and writes back.

  • Embed: If you prefer clicking, you get apps integrated into the chat interface. Canva, Notion, or Asana open in a window beside the conversation.

  • Build in: Providers build what's missing themselves. For Claude, that's now Sheets, Whiteboards, and Data; for ChatGPT, slides and interactive responses.

Visual note: this German graphic shows the shift from switching apps to one chat cockpit, through connecting, embedding, and built-in tools.

From app-hopping to a cockpit: Claude and ChatGPT connect, embed, or build apps

And a quick trip to the stock market

Meanwhile, Anthropic is getting serious about its IPO. Pricing is expected on Wednesday, November 18, with shares trading on Nasdaq under ANT from Thursday. The range is $70 to $78, putting the valuation at around $1 trillion.

Unusually, about 20% of shares are reserved for retail investors, mainly through US brokers. For readers in German-speaking Europe, that means regular trading from Thursday through a broker that offers US stocks. Not investment advice, just so you're aware.

An interesting aside: just before the IPO, Cat Wu, Claude Code's product lead, moved to OpenAI. There, she heads product management for ChatGPT and other products.

OpenAI also seems close to another big model. After Astra, Sol, and Terra arrived at four-week intervals, the codename Betelgeuse, nicknamed Bel, is now circulating. Supposedly an entirely new base model. Names such as "gpt-6.2-betelgeuse" briefly appeared. Nothing official yet. But we're ready. 🍿

Our take: The chat owns the cockpit. You own the rules.

Shopping, sorting emails, planning the holiday party: almost everything in this issue happens through a chat window. Canva is still Canva; you just aren't sitting inside it anymore. Choosing Claude or ChatGPT as your cockpit almost becomes a matter of taste. For app providers, it's a double-edged sword: they're everywhere, but someone else owns the conversation with you.

That makes what you own all the more important. Your fake-store Skill, your inbox operating manual, your files. Custom GPTs show how quickly a provider can shut down something you've invested hours in. You take your rules and data with you. Windows come and go.

And the final click stays yours. For buying as well as sending.

🎬 And finally: Jensen starts over

Barely has Jensen Huang stepped down been fired from Nvidia, and he already has a new job. Not with Elon. And not back at Nvidia, even though over 100,000 people signed for his return. At Dunder Mifflin, at least in this Office clip.

That's a wrap! Your agent is ready to check every Black Friday offer.

Thanks for reading.

Reto & Fabian from AInauten

Login or Subscribe to participate