The Quiet Week Claude Became Your Coworker
Skills, Agents, and the Infrastructure Nobody Noticed
While everyone watched the hype, Anthropic built the foundation for personal AI agents
Wharton professor Ethan Mollick asked Claude to create a PowerPoint presentation for Hamlet. More specifically, the deck McKinsey consultants would present after the whole “ghost of your father” incident.
Claude generated a complete strategy deck right down to the McKinsey Elsinore office branding and a SWOT analysis of Hamlet’s revenge options. Mollick tweeted: “Anthropic has a history of releasing interesting things quietly & without much non-technical explanation. It happened again.”
PowerPoint generation is cute. It’s accessible. It makes for a good tweet.
But it’s not the story.
What I Actually Built This Week
While Mollick experimented with Shakespeare decks, I spent the week building something that would have taken weeks of manual work: a system that documents my AI research automatically and publishes it to my blog without me lifting a finger.
Here’s what happened:
I got a new NVIDIA DGX workstation (essentially a personal supercomputer with 1 petaflop of compute). The first thing I installed? Claude Code.
Then I told Claude what I needed: “Set this up for AI research and make sure everything I do gets documented.”
Claude Code orchestrated the entire setup:
Configured the development environment
Installed AI models and tools
Created a documentation system that captures every experiment
Built a pipeline that syncs my research notes to my MacBook
Automated the transformation of those notes into blog posts
Now, when I run experiments on the DGX, a detailed session file gets created automatically. It captures objectives, code, results, challenges, and learnings. Back on my Mac, I have a simple command that checks what’s new, transforms those sessions into polished articles, and publishes them to my technical blog at publish.obsidian.md/aixplore.
The workflow: Do research → Documentation happens automatically → Blog post generates with one command → Review and publish
Time from “finished experiment” to “published article”: 5-10 minutes
Here’s what makes this interesting: The blog itself (where I publish these articles) was also built and is maintained entirely using Claude Code. Every article, every index, every navigation system, every quality check. All automated.
I’m not working harder. I’m working with an AI that handles the infrastructure while I focus on the interesting parts.
Understanding Skills and Plugins
Before I explain what else I built, it helps to understand what Anthropic released this week alongside Claude Code.
Skills are packaged sets of instructions that Claude loads automatically when relevant. Think of them as expert knowledge you can give Claude once, and it remembers forever.
Instead of explaining “here’s how I want research papers summarized” every time, you create a Skill once. From then on, when you give Claude a paper, it automatically knows your preferred format, what to extract, and how to structure the output.
Anthropic released pre-built Skills for working with PDFs, PowerPoint, Excel, and Word documents. But you can create your own for anything: data analysis, code formatting, report generation, whatever workflows you repeat.
Plugins let Claude connect to external systems and services. They’re how Claude can send emails, access your calendar, interact with cloud services, or sync files between computers.
Together, Skills and Plugins transform Claude from a conversational assistant into an autonomous agent that can execute multi-step workflows, coordinate between systems, and operate continuously, even while you sleep.
That’s what makes the DGX documentation system possible: Skills define how to document sessions, Plugins handle the connection between my DGX and Mac, and Claude Code orchestrates everything.
Then I Started Building Skills
But the infrastructure was just the beginning. So I wrote my first two Skills:
Skill 1: The Research Summarizer
Takes any academic paper as input, extracts key findings and methodology, outputs structured summaries. Now every paper I read gets processed automatically.
For scientists: Imagine your reading list processing itself while you sleep.
Skill 2: The Personal AI Newsletter
Monitors crypto/AI/tech throughout the day, synthesizes developments, writes in my style, uses a mail MCP server to deliver to my inbox at 6 AM every morning.
I wake up to a curated briefing on everything that matters.
I didn’t just automate reading. I automated the entire workflow from monitoring to synthesis to delivery.
What Just Became Possible
Let me be clear about what changed this week:
Before: AI could help you code, draft text, answer questions
After: AI can operate as an autonomous team member, coordinate workflows, test and deploy systems, document your work as you do it, and publish the results. All while you’re offline.
Three releases made this possible:
1. Claude Code - Command-line AI that operates your development environment
2. Plugins - Connect Claude to external systems (mail, cloud, databases)
3. Skills - Packaged expertise Claude loads automatically
My DGX lab notebook system uses all three:
Claude Code orchestrates the documentation and publishing
Plugins handle the sync and file operations
Skills define how to transform sessions into articles
This isn’t “ChatGPT but better.” This is a different category of capability.
Why “Claude Code” Is Actually a Misnomer
Here’s what I realized this week: I barely leave Claude Code anymore.
Not because I’m coding all day. Because Claude Code has become my interface to my computer.
Think about what a computer actually does: it executes instructions. Manages files. Runs processes. Communicates over networks. Everything is just 1s and 0s being orchestrated.
An AI that can write and execute code isn’t just a “coding assistant.”
It’s a general-purpose computer orchestration agent.
Setting up my DGX? Claude Code configured everything: environment, models, networks, documentation systems. Maintaining my blog? Claude Code handles articles, metadata, navigation, quality checks. Building the sync system? Claude Code wrote the scripts, the tracking database, the generation pipeline.
I’m not “using a coding tool.” I’m delegating computer operations to an agent that understands my intent and executes whatever’s needed.
What This Means for You
You don’t need a DGX or a technical blog. But you should understand what just became possible:
For Scientists
Research workflows that document themselves
Literature reviews that process while you sleep
Automated lab notebooks with consistent formatting
Analysis pipelines that run continuously
The paper summarizer Skill I built? 30 minutes to create. No coding required, just clear instructions.
For Business Leaders
Knowledge workers just got force-multiplied
Routine workflows can operate autonomously
Documentation happens automatically during work
Content creation becomes a byproduct of doing the work
My DGX-to-blog system can generate 5+ polished articles per week as a side effect of my research.
The Real Question
Not: “Can AI do my job?”
But: “What am I still doing manually that could document itself and create value while I work?”
The Infrastructure Moment
When the iPhone launched in 2007, tech coverage focused on whether it was better than BlackBerry. The real story was the App Store infrastructure that came 18 months later, creating a platform for millions of developers.
This week feels similar.
The headlines are about PowerPoint generation and PDF creation. Nice features, but not the story.
The real story: The infrastructure for personal AI agents just got built.
Multi-agent coordination (Plugins let agents interact with external systems)
Autonomous operation (Claude Code runs continuously, not just when you’re watching)
Packaged expertise (Skills make specialized knowledge portable and shareable)
Continuous execution (work progresses while you sleep)
I built:
A self-documenting research lab
An automated publishing system
Skills that process papers and deliver briefings
All while my DGX orchestrates itself through Claude Code
So is everyone else who’s paying attention.
What Comes Next
This week, I gave an AI agent the ability to document my research and publish results automatically.
Next week, I’m giving that same agent access to:
1 petaflop of compute
4TB of unified memory
128GB of RAM
An NVIDIA DGX workstation
What happens when you give a general-purpose AI agent access to serious hardware?
Not “run bigger models.” Everyone’s doing that.
I mean: What can an autonomous agent accomplish when compute is no longer the constraint?
Multi-agent systems with dedicated resources
Continuous experimentation at scale
Real-time analysis of massive datasets
Parallel model training and evaluation
Simulations that would take days on cloud infrastructure
I don’t know yet. That’s what makes this interesting.
This isn’t theoretical. The hardware arrived today. Claude Code is already installed and orchestrating the environment and recording the knowledge.
The first blog post from the DGX lab is live now!
Where This Is Heading
The question isn’t whether AI will change how knowledge work gets done.
The question is: Are you experimenting now, or waiting until your competitors figured it out?
Three quiet releases this week:
Skills - Make expertise portable
Plugins - Connect agents to the world
Claude Code - Orchestrate computers autonomously
Together they create something new: Personal AI infrastructure.
Not “AI that answers questions.”
Not “AI that helps you code.”
AI that operates as an autonomous colleague.
The people who figure this out in the next six months (who develop judgment about delegation, who build Skills for their workflows, who learn to work with AI agents) will have capabilities their competitors won’t match for years.
Not because the technology is exclusive. Everything I described costs $20/month and is available to anyone.
But because developing judgment takes experimentation.
That clock started this week.
Next week: What happens when an autonomous agent gets 1 petaflop—first results from the DGX experiments
Follow the DGX Lab Chronicles at publish.obsidian.md/aixplore
I just completed a 5-part technical series on building production-ready ML workspaces on GPU infrastructure, covering everything from workspace organization and documentation systems to experiment tracking, AI agent templates, and team collaboration workflows. What made this possible? Claude Code itself. While working directly on my DGX workstation, Claude Code documented our entire workflow in real-time, capturing the architecture decisions, setup scripts, and best practices we developed. Then, it transformed those raw session notes into a structured 51-minute learning path with interconnected articles, automated all the index updates, and handled the complete publishing pipeline to my Technical AI Articles & Digital Garden blog. The result: practical, battle-tested guidance drawn from actual implementation work, organized into a cohesive series that takes readers from empty directory to full production ML workspace. It’s a perfect example of AI-assisted technical writing where the AI doesn’t just help you write, it helps you capture, organize, and share knowledge as you build.