Hey, it’s Andreas.

In today’s issue:

• Google's $4.75B energy plan

• Anthropic releases Claude Opus 4.6 

• Peking University and Google Cloud AI release PaperBanana

• Super Bowl & AI advertising highlights (2026)

• And more…

Weekly Field Notes

🧰 Industry Updates

🌀 Anthropic releases Claude Opus 4.6 → Anthropic upgrades its top-tier model with stronger coding and reasoning, plus a 1M-token context window.

🌀 OpenAI drops Codex desktop app, Frontier, and GPT-5.3-Codex  → OpenAI released a new desktop Codex app for managing multi-agent workflows, introduced Frontier as an enterprise agent platform, and unveiled GPT-5.3-Codex, its fastest, most general-purpose coding model yet. 

🌀 Elon Musk folds xAI into SpaceX in a $1.25T deal → SpaceX absorbs xAI, implying a ~$1T SpaceX valuation and ~$250B for xAI.

🌀 Google moves into power generation with $4.75B Intersect acquisition → Google is buying Intersect, a wind and solar developer, making it the first major tech firm to directly own large-scale energy production.

🌀 Mistral AI releases Minis 3 model weights → Mistral AI open-sourced 14B, 8B, and 3B models in base, instruction, and reasoning variants, and published its distillation recipe.

🌀 Anthropic spends ~$8M on Super Bowl ads to signal Claude stays ad-free → Anthropic used the world’s most expensive ad slot to differentiate on a single promise: Claude will not run ads.
[Other AI companies' advertisement spots during the Super Bowl:Meta, Amazon, Google, Base44, Genspark.]

🎓 Learning & Upskilling

📘 LandingAI launches Document AI course on agentic extraction → LandingAI released a short course teaching how to build agentic pipelines that turn PDFs into structured Markdown and JSON.

📘 Stanford Online hosts webinar on human-centered AI design → Stanford Online explores how to design trustworthy AI systems, with Stanford HAI leaders outlining human-centered principles.

🌱 Perspectives & Research

🔹 Peking University and Google Cloud AI release PaperBanana → Five-agent system auto-generates publication-ready diagrams and charts for academic papers.

🔹 Elon Musk, Dwarkesh Patel, and John Collison podcast → Nearly 3-hour deep dive on orbital data centers, China’s AI trajectory, power constraints, and the roadmap for xAI and Grok. Wide-ranging, technical, and very forward-looking.

🔹 OpenClaw creator Peter Steinberger hosted by Y Combinator on why personal AI agents could kill most apps → Arguing that local-first personal agents will absorb app functionality, making up to 80% of today’s apps obsolete.

🔹 IBM Technology covers security and governance for autonomous AI agents → IBM Technology breaks down real agent risks - prompt injection, data poisoning, bias - and the safeguards needed to build secure agentic systems.

♾️ Thought Loop
What I've been thinking, building, circling this week

Three weeks ago, Anthropic released Claude Cowork - essentially Claude Code for non-technical users. An agentic AI that runs on your Mac, reads your local files, and handles multi-step tasks without hand-holding. It was interesting. Then, one week later, they added plugins.

That's when Wall Street panicked.

The release of Anthropic Cowork Plugins triggered a one-day selloff that wiped roughly $285B in market value across software, data, and adjacent sectors. Traders coined a name for it: "SaaSpocalypse." Thomson Reuters fell nearly 18% in a single session. RELX dropped about 14%. Wolters Kluwer shed roughly 13%. Investors repriced anything that looked exposed to AI-driven workflow automation.

The main reasons were:

• Markets internalized that Claude Code and AI agents can automate large parts of SaaS-style workflows (legal, sales, finance, ops) directly via files and plugins, bypassing many incumbent software platforms.

• This triggered fear-driven repricing: if agents compress demand for traditional SaaS, outsourcing, and IT services, revenue multiples break.

Which makes this a good moment to slow down and actually look under the hood.

Below is everything you need to know about Claude Cowork plugins - how to get started, how to customize them for real workflows, and where their limits still are.

What is a Claude Cowork plugin?

First, the foundation. Claude Cowork is an agentic AI assistant that runs inside the Claude Desktop app on macOS. It's built on the same engine as Claude Code, Anthropic's developer tool. But instead of targeting engineers, Cowork targets knowledge workers - people who spend their days in documents, spreadsheets, email, and research.

The key difference versus a standard chatbot is simple: Cowork runs on your machine. It can read and write local files, browse the web, execute code, and chain multi-step tasks autonomously. This is less “answer my question” and more “take ownership of the project.” Crucially, it also works in the background. While Cowork is researching, drafting, or building, you can continue using Claude for other tasks. The agent keeps accumulating context and progressing independently, which is what turns it from a chat interface into a genuinely powerful workflow engine.

So Claude Cowork is basically an AI that can read your files, run code, and handle multi-step tasks without hand-holding. But out of the box, it's a generalist. It knows a little about everything and a lot about nothing specific to your job.

That’s where Claude Cowork plugins come in.

A Claude Cowork plugin turns this generalist agent into a role-specific operator. Each plugin bundles three core elements into a reusable setup, shaping how Claude works, what it can do, and how it executes tasks. Today, there are 11 official plugins available, all open-source.

What’s inside a Claude Cowork plugin?

The answer is refreshingly simple.

Claude Cowork plugins are essentially Markdown files. They contain step-by-step workflow instructions Claude can execute autonomously, triggered via slash commands and optionally extended with connectors - bridges to other software via MCP.

There’s no complex framework and no proprietary format. If you can write Markdown, you can build a plugin.

Here's the basic anatomy of a plugin's file structure:

plugin-name/
├── .claude-plugin/plugin.json   # Manifest (JSON: name, version, description)
├── .mcp.json                    # Tool connections (MCP servers)
├── commands/                    # Slash commands as Markdown
│   └── call-prep.md
└── skills/                      # Domain knowledge as Markdown
    └── SKILL.md

Anthropic found a clever way to package domain-specific prompting as portable, shareable file bundles. These plugins are transparent. Every instruction the AI follows lives in readable text files. You can open them, audit them, and change them. No black box. What's behind it is essentially structured prompt engineering in file form.

How to install and customize a Claude Cowork plugin

Two prerequisites. You need a paid Claude subscription — Pro, Max, Team, or Enterprise all work. And you need the Claude Desktop app for macOS.

Open the app on your desktop, and you will see a "Cowork" button in the top center.

After opening Claude Cowork, click the plus icon on the left bottom and expand the section. From there, search for the plugin you want to work with.

Once selected and installed, Claude immediately exposes the relevant slash commands and built-in skills you can start using.

From my experience, customize before you go any further.

The out-of-the-box plugins are built for a generic version of your role. If you skip customization, results will feel underwhelming and shallow. The real leverage comes from tailoring them to your company, your processes, and your tools.

Unfortunately, this step is currently not obvious and feels counterintuitive in how Claude Cowork Plugins are set up. Customization is optional, buried, and easy to overlook. I hope this changes - ideally making customization clearer and mandatory, because that is where most of the value actually sits.

Where the real value lives: CUSTOMIZATION

Customization ensures the plugins operate with the right structure, behavior, and expectations for how you actually work - not how a generic demo assumes you do.

Because plugins are just files, customization is simply editing text.

After installing a plugin, go to the top-right corner and click Customize. From there, Claude Cowork Plugins will guide you through the full setup by asking structured questions using the AskUserQuestion tool.

Real-world Claude Cowork plugin use cases

Overall, Anthropic released 11 open-source plugins. Here's what each one actually does:

The reality check

Cowork plugins are powerful after customization. The concept is genuinely interesting - Anthropic found a clever way to package domain-specific prompting as portable, shareable file bundles. Open-source. Transparent. A level of control that's uncommon in AI tooling.

But there's a gap between the announcement and the experience.

On paper, this sounds accessible. In practice, it comes with real friction. Even simple tasks - inputting data into a spreadsheet, for instance - feel clunky and time-consuming. Setting up a plugin means configuring markdown files, wiring MCP connections, and troubleshooting integrations with nearly no documentation to guide you. Anthropic didn't exactly roll this out with a keynote. They dropped it via a single tweet and labeled it a research preview. The sparse docs reflect that.

Most knowledge workers won't do such a clunky setup. They want tools that work off the shelf. They're not going to edit YAML frontmatter or debug a .mcp.json file to connect their CRM. The people who will - technical early adopters, power users, developers building for their teams - are a narrow slice of the workforce.

And then there are the practical constraints. Cowork is macOS-only. Sessions are local and don't sync across devices. There's no native GSuite support - you can access Google Drive through desktop sync clients, but that's a workaround, not an integration. And Cowork tasks burn through your usage allocation fast. A few multi-step workflows can eat 15% of your token budget in minutes. The Pro plan at $20/month will feel restrictive for daily use. The Max plan at $100/month is the realistic choice for anyone doing serious work.

So is this a threat to established SaaS companies? Does it explain a $285B market crash?

Not today. The SaaSpocalypse was definitely an overreaction to a research preview that most knowledge workers will never configure. The plugin system is a smart proof of concept - a signal of where agentic AI might heading - but it's not replacing Salesforce, Zendesk, or Thomson Reuters anytime soon. The companies that lost 13–18% in a day aren't competing with markdown files on someone's MacBook. They're competing with whatever this probably becomes in two years, once it's polished, cross-platform, cloud-native, and zero-config.

For now?

It's worth trying. If you're the kind of person who enjoys tinkering with AI workflows and wants a transparent, customizable assistant on your own machine. It won't change the world tomorrow. But it’s a glimpse where we might going next with knowledge work.