SUBSCRIBE | ADVERTISE

I tested Claude Cowork

AI news sans hype from the last 10 days + my two cents on Claude Cowork

Was this email forwarded to you? Sign up here.

Howdy wizards,

New tools & product features

• Anthropic released Claude Cowork, essentially a simpler version of Claude Code that’s meant for non-developers. Just like Claude Code, it can manage local files on your computer, complex tasks, and run parallel workflows. It was actually built by the Anthropic team in just 10 days, using mostly Claude Code itself. It’s available on Claude’s paid plans and only inside the Claude MacOS app. Read my take on Claude Cowork later on in this email ↓

• OpenAI announced they’ll begin testing ads in ChatGPT for free and Go tier users in the US. Ads will look like sponsored recommendations based on users’ conversations, excluding some areas like health and politics. Sam Altman called ads "a last resort" late last year, but let's be real: OpenAI made $20B in 2025, while Meta pulled $180B and Google $295B purely from ads. This was never about reluctancy to ads; it was about timing. Build the habit first, monetize second. Looks like the ChatGPT enshittification era is upon us.

• Google launched Personal Intelligence, letting Gemini reason across Gmail, Photos, YouTube, and Search data to give you personalized responses. Google has a massive competitive advantage here that its competitors can’t match: decades of your searches, emails, docs, locations, and preferences.

Industry moves

• Apple and Google have signed a multi-year partnership, with Gemini becoming the backbone of Apple's AI strategy, including the long-awaited Siri upgrade. Apple has been lagging on AI for years, and seems to now have decided they’re not competing head-on in foundation models. Big win for Google here, who’s now the world's 2nd most valuable company.

• Anthropic cut off xAI's access to Claude Code after discovering they were using it through Cursor for internal development. Anthropic's terms explicitly ban using its models to build competing AI systems. xAI’s cofounder told staff they'll take a hit on productivity, but that it will also accelerate work on their own coding tools. This is part of a broader crackdown from Anthropic — they've also blocked OpenCode recently, an open-source alternative to Claude Code that previously let users connect their Claude accounts directly. Claude Code is having a ChatGPT moment right now, and Anthropic wants to protect it.

• xAI's Colossus 2 supercomputer is now online—the first gigawatt-scale AI training cluster in the world. xAI has been building infrastructure faster than anyone else; will be interesting to see if it actually translates into more powerful AI, too.

IN PARTNERSHIP WITH AUTH0

Authentication for AI agents, integrated in minutes

Building secure authentication for AI agents from scratch takes tons of development. Handling user identification, permission enforcement, secure app connections, and approval flows. Auth0 for AI Agents gives you all of this, production-ready, and you can integrate it in minutes instead of months.

Auth0 has been doing identity at scale for years (10 billion+ authentications every month). Now they've built it specifically for AI agents, so you can focus on what makes your product unique instead of reinventing auth infrastructure.

Start building today →

My thoughts on Claude Cowork

I’ve been messing around with the new kid on the block—Claude Cowork, for a few hours.

First of all—it’s fantastic that there’s now a simpler alternative to Claude Code out there. It will get more people used to the power of AI that can actually do things/work agentically.

I saw a Redditor saying that Cowork was a top 3 most exciting tech moments for him, likening it to getting a Gameboy for Christmas at age 9. The stakes are monumental, folks!

He then went on to create his own version of Wispr Flow, an AI voice dictation app*, by having Cowork use Claude Code to create it (total inception!). A non-engineer being able to just spin up a product like this, even though it’s probably pretty buggy, is amazing.

The creator of NodeJS tweeted this week that “the era of humans writing code is over” and I don’t think he’s wrong.

In my testing of Claude Cowork, the first thing I did was asking it to reorganize my Downloads folder; a monstrosity which had blown up to reach 13,478 files! It took several back-and-forth prompts to get right, but in the end it managed to organize everything neatly. It definitely took extra time because of the amount of files.

Cowork feels useful to me. However, as a Claude Code user for nearly a year now, it also feels like a limiting wrapper around the real thing. I’ve gotten totally accustomed to CC at this point. I probably won’t be using it unless they launch some feature that makes specific workflows that I need way more convenient inside Claude Cowork.

So, who do I think Cowork is for?

If you have a lot of meetings and a busy calendar and inbox, I imagine it could be a good surface to help manage your workday and organise your schedule and such. Personally, I’m not going to use it for that because as much as I enjoy using AI for so many things, I love organizing all my to-do’s with pen & paper.

Also, if you have been curious about trying Claude Code for any reason at all, but it sounds scary to you to move away from the familiar Claude interface, then I definitely recommend testing Cowork. It’s likely a gateway drug!

A note on Cowork and security

One thing to keep in mind if you haven’t used an AI tool with access to your computer before: security concerns are real. Prompt injection attacks could become more serious since the agent has access to your computer, such as leaking your data, or deleting it. Not saying this to scare you, but to proceed with caution.

Here’s an example of how a prompt injection hidden in a word doc that the user uploads could turn into data leakage.

But even without malicious actors, you can shoot yourself in the foot. For example, I asked Cowork to help me clean up duplicates in the folder, and nearly accidentally deleted some unique files.

Here’s what happened: I asked Claude for a duplicate cleanup. Then, to make sure it was about to do this safely, I asked what method it intends to use for identifying duplicates (to know if it’s checking the actual content). It told me it wanted to use filename-based patterns to remove duplicates! That means if I have two files both titled img_2.png, then those would be listed as candidates for deletion even though they might be entirely different images.

On the bright side, being curious enough to ask clarifying questions like this before letting the AI proceed with an action, actually goes a long way in reducing risk.

In this case, I told Cowork to use content-based hashing instead (I didn’t know about it before, but it explained alternative methods to me), which actually compares the content and is safe. This allowed me to quickly free up several GB of files.

❦

*Btw. Are you still writing to AI instead of talking? Using a dictation app to transcribe my words in real-time saves me at least a couple of hours every day. If that sounds interesting (and you don’t feel like coding one yourself), I warmly recommend getting Paraspeech for Mac. It’s the only one in the market with a cheap lifetime license. It’s also blazing fast and doesn’t collect your data.

THAT’S THE LETTER.

Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.