Hey, it's Andreas.
A quick heads-up: last week I soft-launched The Agentic AI Cohort, just to people who'd already raised their hand, to see if there was real demand or only polite interest. Instead, 17 of the 20 founding seats were gone in three days, before I've even opened it to the public.

But the part I'm most excited about is the mix of people who got in: VPs, directors & project leads from companies you'd recognize, people who've decided to stop watching AI happen and start running agents that do the work for them (and hand a few hours a week back in the process).

I'll share more this week and launch the cohort, so keep your eyes open. Three founding seats are left, they'll probably go today, and then we open it to the public.

On to the news.

In today's issue:

Let's get into it.

🌀 Anthropic’s Claude Fable 5 gets pulled after three days → Anthropic launched Fable 5 as its most capable general-use model, then disabled it after a U.S. government export-control order. Frontier models now look less like apps and more like critical infrastructure - powerful, metered, and suddenly revocable.

🌀 Jeff Bezos’ Prometheus raises $12B for physical-world AI → Prometheus is building what Bezos calls an “artificial general engineer” for complex machines like jet engines and medical devices.

🌀 Apple rebuilds Siri around Gemini-powered AI → At WWDC, Apple introduced Siri AI, routing requests across on-device models, Private Cloud Compute, and Gemini-based cloud models.

🌀 SpaceX becomes the largest IPO in history → SpaceX began trading under SPCX after raising about $75B at a valuation near $1.77T. The real story is probably not just rockets - it is Starlink cash flow, xAI exposure, and public-market access to Musk’s AI + space infrastructure stack.

🌀 OpenRouter launches Fusion for multi-model answers → Fusion sends one prompt to multiple models, ranks the outputs, and merges them into a final answer. The pitch is timely: frontier-level performance may come less from one dominant model and more from cheaper model panels stitched together through an API.

🌀 China’s universities cut 12K programs as AI reshapes jobs → Chinese universities have scrapped more than 12,000 programs over five years, shifting away from arts and languages toward AI, engineering, and tech fields.

🌀 Former xAI co-founder launches River AI for personalized agents
→ Igor Babushkin’s new startup is building agents that adapt to each user’s style, goals, and workflows.

📘 Alteryx shares Inspire 2026 on-demand sessions → A hands-on way to catch up on the latest Alteryx updates, CEO keynote, and hands-on platform demos. Best starting point: Josh Burkhow’s demo around the 19-minute mark for a practical look at how AI-powered analytics workflows are evolving.

📘 Stanford puts CS153 Frontier Systems online → Stanford’s new lecture series is now available on demand, featuring frontier-system builders like Jensen Huang, Ben Horowitz, and others. Brilliant watchlist for anyone tracking AI-native companies, resilience, infrastructure, and the next operating layer of frontier tech.

📘 Developers Digest shares a 7-minute Codex primer → Quick overview of Codex Desktop, including Plan Mode, Goal Mode, Plugins, multi-agent workflows, worktrees, browser annotations, and scheduled automations. Good quick start!

📘 DeepLearning launches course on fast LLM inference with vLLM → A practical short course with Red Hat on deploying open-source LLMs efficiently.

📘 Zack Proser (WorkOS) on why attention is the real agent bottleneck → A practical AI Engineer talk on working with parallel agents without burning out. Covers voice-first dispatch, isolated worktrees, verification gates, signal agents for Slack/Linear, and self-improving workflows built from your own agent history.

📘 Anthropic reflects on one year of Claude Code → Some good insights how Claude Code has moved from internal terminal tool to enterprise coding workflow.

🔹 Matt van Horn and Addy Osmani on designing agent loops →There's been significant buzz around loops and loop engineering lately. To quickly get up to speed on the conversation, here are two insightful reads. In short, the future of agentic work isn't about crafting better prompts; it's about developing more effective loops.

🔹 Satya Nadella argues companies need token capital → Nadella says AI advantage will come from owned learning loops, workflows, context, and human judgment - not just access to frontier models. He says the moat is not the model alone, it is how organizational knowledge compounds inside the system.

🔹 Dario Amodei warns that AI policy is moving too slowly → Anthropic’s CEO argues for an FAA-style AI safety regime with mandatory third-party testing and government deployment powers.

🔹 Perplexity & HBS study how agents change knowledge work → Perplexity compared 10k tasks across Search and Computer, finding agents pushed users toward harder, more creative, cross-domain work.

🔹 Bloomberg goes inside Anthropic’s rise → Emily Chang sits down with Dario and Daniela Amodei for a look at Anthropic’s origin, safety-first positioning, and growing tension with government and defense work.

🔹 OpenAI shows how a Japanese farmer uses ChatGPT and Codex → Hiroki Tomiyasu uses ChatGPT and Codex to diagnose crop issues, monitor fields with satellite data, control greenhouse motors, build LINE bots, and design farm databases. Unconventional real-world example of AI as an always-available technical partner for non-technical operators.

Here is the simplest way to 10x your AI results: stop writing longer prompts, and start building reusable context files.

I've watched thousands of people use AI. I can usually tell within a few minutes who gets great output and who gets noise. It comes down to one habit, and almost nobody has it.

Smart, senior, technical people included. They open ChatGPT or Claude, type a request the way you'd ask a stranger on the street, then judge the model by the generic answer it hands back.

The model isn't the problem. It has no idea who it's talking to. So it gives you the average. Average advice, average writing, average strategy. Good enough for a demo, too generic to trust on real work.

Think about onboarding a brilliant new hire and telling them nothing. Not what the company does, who the customers are, what your goals are, or how you like things done. Day one they're sharp, fast, and useless. Because they're guessing.

The fix isn't a better prompt

It's well-written context files, written once. You brief the model the same way you'd brief that new hire, then reuse it.

But not your whole life in one folder. This is where most people overcorrect. They connect every note, every doc, every old email, and the output gets worse, not better. This quickly leads to what is known as context rot. The model starts citing a goal you abandoned two years ago and a project that shipped last spring, with total confidence. Too much context is its own problem: the model drowns in the noise and goes back to guessing.

So the failure has two faces. No context, and the model guesses. Too much, and it drowns. Both land you in the same place: a generic answer that doesn't fit your situation. The skill is picking the few things that actually change the answer, and keeping them current.

You only need five files

I spent hundreds of hours testing different AI setups. At first, I had a massive stack of markdown files. Then I stripped everything back to almost nothing. Then I tried using just one master file. But after all that testing, I kept coming back to the same conclusion. You do not need a huge context repository. You need five short files. That is the setup I now trust because it is simple, practical, and proven.

→ who-i-am: so it stops treating you like a stranger
→ working-with-me: how you want it to behave
→ goals: what you're optimizing for, and what to ignore
→ context: your work, your customer, your real bottlenecks
→ voice: so it writes like you, not like an AI

The trick is loading them by task or project, not by default. A client proposal pulls who-i-am, goals, and context. Writing a LinkedIin post or Newsletter pulls voice. You bring the slices the task in front of you needs, the same way a good new hire doesn't reread the whole company wiki to answer one question.

The voice file earns its place

One of these does something the others don't. AI writing has a tell now. The same flat, agreeable register. The same words: leverage, delve, unlock, robust, seamless. The em-dash wedged into every other sentence. The tidy rule of three. Readers spot it in about two lines, and the moment they do, your work reads like everyone else's.

You don't fix that with a "write better" instruction or “write like a human would”. You fix it by handing the model your actual voice, pulled from things you actually wrote, plus a short list of the tics it should never use. That is what the voice file does. The model stops reaching for the average sentence and starts reaching for yours.

What this means

I packaged the exact setup I use into a free kit: The Context Repository.

It includes five short files, plus the prompts that help you build them. Each file works like a guided interview. You paste in the prompt, answer the questions, and save the final document.

It will take you around 30 minutes to 1 hour to complete properly. So don’t rush it. Don’t open it between meetings. Block the time.

Because if you do this well, it will become one of the highest-leverage upgrades you can make to your AI output. You also don’t need to use all five files in every chat or project. That would be overkill. Use the files that match the task.

Because if you do this well, it becomes one of the highest-leverage upgrades you can make to your AI output.

But here is the important part: The files only work if your AI can actually access them.

Don’t just create the files and leave them sitting somewhere on your laptop. Your AI needs to see them, read them, and use them before it starts working.

You have two options.

Whenever you start a new chat, upload or paste the files that matter for that task. Then tell the AI to read them first.

You do not need all five files every time. That would create context bloat. Use the files that match the task. For example, if you are writing a newsletter (like I am doing right now), you may need your voice, who-am-i, and context files.

The better setup is to save all five files in one dedicated project or workspace. For example, upload them into a Claude project, Claude workspace, or ChatGPT Project if your tool supports it.

Then, whenever you start a new serious task, refer the AI back to those project files and tell it to use the relevant ones before answering.

The key is this: Always give the AI context first and refer to it.

Don’t just upload files and assume the model will use them correctly. Tell it exactly what to read, what to ignore, and how to apply the context.