Hey, it’s Andreas.
I’ve received a lot of feedback great feedback on this newsletter, and one message came through clearly: you want more real-world AI applications, especially practical ways to use AI agents in everyday life for you personally.

So from now on, the deep dive section will be exclusively actionable and focused on doing, not just discussing. I still care about the bigger conversations around AI, jobs, and long-term impact. But your feedback made one thing clear: what’s most useful right now is practical guidance you can apply immediately.

That’s why I’m working on a comprehensive guide to how I use AI in my daily life - from writing this newsletter to coding, building a personal brand, and getting real work done faster.

I may also turn it into a video walkthrough.

In today’s issue:

Let’s get into it.

🌀 Figure AI founder Brett Adcock launches Hark → Hark is a new AI lab backed by $100 million of his own capital and aimed at building what he calls the world’s most advanced personal intelligence.

🌀 Anthropic’s leaked Mythos materials hint at a new model tier → Leaked launch materials suggest Anthropic is preparing Claude Mythos, a model above Opus, while the company has confirmed it is testing a new general-purpose system with major advances in reasoning, coding, and cybersecurity.

🌀 OpenAI shuts down Sora to refocus on core bets → OpenAI is reportedly shutting down both the Sora app and API as it cuts side projects and reallocates compute toward its next flagship model, codenamed Spud.

🌀 ElevenLabs and IBM bring voice to watsonx Orchestrate → ElevenLabs has integrated speech-to-text and text-to-speech into IBM watsonx Orchestrate, giving enterprise agents more natural multilingual voice capabilities across workflows.

🌀 Google Research introduces TurboQuant → Google Research unveiled TurboQuant, a compression method that cuts model memory use by more than 6x without retraining while delivering up to 8x speed gains on Nvidia H100s with almost no accuracy loss.

🌀 Meta open-sources TRIBE v2 for brain activity prediction → Meta has released TRIBE v2, a model trained on 500+ hours of fMRI data from 700+ people that predicts how the brain responds to sights and sounds across tasks, subjects, and languages.

📘 Deeplearning.ai x Anthropic course on Agent Skills → A beginner-friendly short course on how to build reusable skills for agents using standard formats, on-demand loading, and modular workflows.

📘 Anthropic guide to Claude Skills → Anthropic’s official 33-page guide on Claude Skills is still one of the best practical resources for learning how to turn repeatable prompting into reusable workflows.

📘 Tutorial Microsoft VP on automating daily work with micro-agents in Warp
→ A practical tutorial showing how Microsoft’s AI VP uses Warp to automate repetitive tasks with lightweight micro-agents, from managing Azure resources to processing files and streamlining workflows with custom rules and prompts.

📘 A practical roadmap for getting started with AI agents in 2026
→ If you want to move from playing with AI tools to deploying real agents, the path is simpler than most people make it sound.

🔹 WSJ on the Altman-Amodei rivalry shaping the AI race → The Wall Street Journal traces the decade-long fallout between Sam Altman and Dario Amodei from their OpenAI days to today’s OpenAI-Anthropic rivalry, arguing that personal conflict, broken trust, and competing visions now shape two of the most important companies in AI. Useful reminder that frontier competition is not just about models and capital, but also about the people steering them.

🔹 Andrej Karpathy on the LiteLLM supply chain attack → A compromised LiteLLM release on PyPI exposed users to credential theft, hitting one of the most widely used packages in the AI stack and forcing emergency remediation for affected versions 1.82.7 and 1.82.8. Good breakdown of what happened.

🔹McKinsey & Company on AI trust maturity in 2026
→ McKinsey finds AI trust maturity is improving, with average RAI maturity rising from 2.0 to 2.3, but only one-third of organizations score 3 or higher.

🔹 IBM and Palo Alto Networks on securing AI-driven operations
→ In a new report based on 1,000 C-level executives, IBM and Palo Alto Networks examine how enterprises are protecting AI-powered operations and where AI itself is starting to strengthen security.

🔹 Stripe on how its engineering team built coding agents → Stripe shows how internal AI agents, or “minions,” are triggered from Slack to write code, deploy changes, and even make machine-to-machine payments to buy services or call external APIs.

🔹 Intercom on the rise of vertical AI models → Intercom says its customer service agent Fin now outperforms top general models like GPT-5.4 and Opus 4.5 on its domain, while staying fast, cheap, and highly scalable.

🔹 Mario Zechner on why builders should slow down with coding agents
→ In a recent post, Pi creator Mario Zechner argues that software quality is slipping as teams lean harder on coding agents without enough human judgment, review, or restraint.

Claude Skills are a major unlock for agentic workflows because they turn repeated expertise into reusable, on-demand context instead of brittle one-off prompts. They also help agents act more consistently by packaging instructions, resources, and even executable steps into a structure Claude can load only when needed.

I’ve spent a lot of time writing about this in the past, and I’ve received many questions from readers and clients asking how to write Skills well, how to improve them, or why they struggle to set them up in the first place.

Over the last few weeks alone, I’ve created more than 170 Skills. One of the most effective things I’ve started doing is using a meta approach: I ask Claude Code to build the Skill for me.

Let me explain what I mean.

But first of all a short recap: A skill is a note you leave for Claude. Written in plain English. Saved as a file. That's it.

It's not code. It's not a plugin. It's not an API integration. It's just writing clear instructions in plain English, saved in a structured folder.

Skills are instruction packages that give Claude specialized knowledge it does not have built in. Claude knows a lot. But it does not know your company's brand guidelines or how you write your newsletter. It does not know the specific way your team formats financial reports. It does not know the workflow your industry uses for compliance documentation.

Skills fill those gaps. They provide four types of knowledge:

1 ➜ Step-by-step workflows that tell Claude exactly how to complete a process in order.

2 ➜ Domain expertise that gives Claude the rules and standards for your specific field. E.g. what is acceptable in healthcare documentation. Or how legal contracts should be structured.

3 ➜ Tool integrations that teach Claude how to work with specific file formats the right way. Not just creating an Excel file, but creating one with working formulas and proper formatting.

4 ➜ Reusable resources including scripts, templates, and reference docs that Claude can pull from when needed. Your actual templates, not generic ones.

One critical distinction: skills are not just fancy prompts. They are structured packages that persist across conversations and include actual files Claude can work with.

What makes Skills powerful is not the file itself. It is what you put inside it.

Most people understand the theory of context.
Yes, agents need the right information.
Yes, better instructions improve outputs.

But that still leaves the actual problem:

How do you get Claude to write the way you write?
How do you make it understand your standards, your taste, your structure, your tone, your examples, your red lines?

That is where most people get lazy.

They throw in a few style notes.
"Write clearly."
"Be punchy."
"Sound like me."

That is not enough and has many worked well in 2023.

If you want a good Skill, you need to extract the stuff that is still trapped in your head.

A simple pattern that works very well is this:

Get Claude to interview you.

Tell it to ask you one question at a time until it has enough material to build the Skill.

There are two ways to get this started. Very open and less specific.

Alternatively, you could approach the task in a more comprehensive way.

That is a much better starting point than just a blank page or writing from scratch. Because now Claude is not guessing.

It is helping you surface the exact context it will later need to perform well.

You can read through the created Skill and keep iterating if you don’t like certain outputs.

P.S.: I wrote about how to set up skills correctly and integrate them into your workflow here.

P.S.S.: If you want to read more detailed information about skills, I recommend this guide.

P.S.S.S.: Someone consolidated all the recommendations from the Head of Claude into a single skill. It's worth taking a look, especially if you use Claude Code.