Claude Cowork Might Be the Most Consequential Piece of Software You're Not Using

I built 4 plugins, 39 skills, a genomic analysis pipeline, and a full-stack app — all inside a desktop agent that most people treat like a chatbot. Here's the operating manual.

TL;DR: Claude Cowork is Anthropic’s desktop agent — a sandboxed VM on your computer that can read your files, run code, control your browser, send messages, and talk to external services. Most people open it, ask a question, and close it. That’s the vending machine approach, and it wastes 90% of the surface area. Over the past ~ two months, I’ve used Cowork as my primary working environment for everything from open-sourcing a family assistant to decoding my own DNA. It’s not a chatbot. It’s a workstation. And I think it’s the most consequential piece of software — ironically, coded largely by AI itself — to ship in the past decade.

Fair warning: A friend read an early draft and told me he couldn't open it without immediately wanting to build fifteen things. Don't read this on a phone between meetings. Set aside an hour, open your laptop, and have Cowork ready. You're going to want to follow along.

---

Why This Matters Now

Reid Hoffman recently wrote that we’re all becoming gamers — not in the console-and-controller sense, but in the tool-building sense. His argument: for centuries, humans built tools to get ahead. Then economies matured, and most of us stopped building tools and started relying on the ones already available. AI reverses that. Platforms that let you turn ideas into real things — instantly get unstuck, operate beyond what one person could normally do alone — make life feel like leveling up in a game. Each challenge is a level. AI is how you craft your way through.

He’s describing something I’ve been living inside of. And I think Cowork is the clearest embodiment of that shift.

Ethan Mollick’s recent piece on management as AI superpower lands the complementary insight: the skill that matters with AI agents isn’t prompting. It’s delegation. His MBA students — doctors, managers, company leaders, most of whom had never coded — built functional startup prototypes in four days using Claude Code. They succeeded not because they were AI experts. They succeeded because they knew how to scope problems, define what “done” looks like, and recognize when output was off. Management skills became their AI skills.

And Anthropic’s own Economic Index confirms the pattern at scale: 52% of users now work with AI as a thinking partner rather than just delegating for full automation. Tasks requiring college-level skills were sped up by a factor of 12. The shift isn’t “AI does the work.” It’s “you set intent, constraints, and a definition of done — then review.” Classic management, applied to software that can improvise.

Nate Jones captures the practitioner angle: AI agents amplify top human talent, making them “more fingertippy.” Agents aren’t future tech — they’re production infrastructure now.

Translation: the gap between “AI as chatbot” and “AI as operating system for your work” is closing fast. And Cowork is where that gap collapses for non-developers.

---

What Cowork Actually Is (And Isn’t)

Cowork runs a lightweight Linux VM on your computer. You select a folder from your filesystem, and the agent can see everything in it. It writes code. It runs scripts. It creates documents. It fills PDFs. It reads your email. It controls your browser. It sends iMessages.

It’s not a chat window. It’s a sandboxed computer that an AI operates while you watch, direct, and approve.

But out of the box, it’s a capable amnesiac. It doesn’t know your name, your insurance carrier, your Slack channels, or what voice to use when drafting a LinkedIn post. Like any new hire, it needs onboarding. Unlike any new hire, it reads the onboarding docs in 200 milliseconds and never forgets them.

The rest of this piece is the onboarding manual I wish someone had given me on day one.

---

Layer 1: Extensions — The Nervous System

Before you build anything, you need to wire the agent to your world. Extensions are Cowork’s MCP connectors — they let Claude interact directly with apps, data, and services on your machine. Think of them as the nervous system: without them, the agent can think but can’t feel or reach.

Here’s my stack:

Desktop Commander — The single most important extension. Terminal access, file operations, process management. This is what lets Claude actually do things on your machine instead of just talking about them. Run git commands, execute Python scripts, manage background processes. Without this, Cowork is a chatbot with a nice UI.

Control Chrome — Browser automation. Claude can navigate to URLs, read page content, execute JavaScript, interact with web apps. This is how it fills out web forms, checks dashboards, researches topics, and interacts with any SaaS tool you use in a browser.

Filesystem — Direct file access beyond the mounted folder. Read, write, search, and organize files across your system.

Read and Send iMessages — Claude can read incoming messages and draft replies. Useful for family coordination, quick responses, and staying in loop without context-switching.

Read and Write Apple Notes — Access to your notes app. Claude can search, read, create, and update notes. If you keep reference material in Notes, this is how the agent accesses it.

Control your Mac — AppleScript execution. This is the wildcard — it lets Claude automate virtually anything on macOS. Open apps, move windows, trigger Shortcuts, interact with system-level features.

PDF Tools — Read, fill, extract, compare, and create PDFs. This is what makes form-filling work — Continuity of Care forms, insurance claims, school registration packets, government paperwork.

The extensions matter because they define the agent’s surface area. Every extension you add expands what “help me with this” can mean. An agent with Desktop Commander, Chrome, and PDF Tools can do things that would take a human twenty minutes of tab-switching and copy-pasting. An agent without them is a fancy autocomplete.

---

Layer 2: CLAUDE.md — The Operating Manual

Every session, Claude reads a file called CLAUDE.md from your workspace root. Think of it as persistent onboarding documentation — except the new hire reads it in 200 milliseconds and never forgets it.

Mine contains my working philosophy, folder structure, preferences (”run terminal commands directly — don’t ask me to copy-paste”), key reference paths, quick facts (family members, insurance IDs), and behavioral instructions.

This isn’t a prompt. It’s an operating manual. And it compounds. Every time I add a preference or a reference path, every future session gets slightly smarter. The model didn’t get better. The context did.

---

Layer 3: Skills — Persistent, Modular Knowledge

Claude skills are folders containing a markdown instruction file and optional reference data. Claude loads them on-demand when it recognizes a relevant task. They’re persistent, modular, and composable — everything a prompt isn’t.

I’ve built 39 skills. Some are simple (a voice and style guide that defines my cadence, vocabulary, and anti-patterns). Some are complex (a family assistant with six reference files covering insurance, medical, financial, household, preferences, and family member data).

Here’s the concrete version. When I say “fill out David’s camp health form,” Claude pulls his DOB, allergies, medications, insurance details, and pediatrician contact from the family assistant skill. It flags his tree nut allergy. It asks me to confirm before submitting. Thirty seconds instead of twelve minutes of hunting through filing cabinets and email threads.

The whole thing is structured markdown and a well-written instruction file. Nothing to deploy, nothing to maintain. I open-sourced the template: github.com/ericporres/family-assistant-skill

The distance between “AI assistant” and actual assistant is shorter than you think — and it’s paved with markdown, not code.

---

Layer 4: Plugins — Workflows That Compound

Skills give Claude knowledge. Plugins give Claude workflows.

Cowork ships with a growing marketplace of first-party plugins from Anthropic — Customer Support, Data, Finance, Legal, Marketing, Product Management, Productivity, Sales, and Bio Research. Each one bundles domain-specific skills with slash commands and workflow templates. They’re one-click installs and instantly expand what the agent can do.

But the real power is building your own. A Cowork plugin bundles skills with MCP connectors and custom commands. Once installed, it activates with a slash command. The anatomy is straightforward: a plugin.yaml manifest, a skills/ folder with your SKILL.md and reference files, and a commands/ folder that maps slash commands to skill triggers.

I’ve built three personal plugins that I use every morning:

Email triage (/email) — Connects to Gmail via the Gmail MCP connector, scans the last 24 hours, classifies every message into three tiers: reply needed, review, and noise. Drafts replies for the urgent ones in my voice. The secret weapon is alias routing — I run 300+ email aliases (banking@, school@, shopping@), and the plugin maps each alias to a default priority tier. Claude classifies 80% of my inbox without reading a single email body. The remaining 20% — personal emails from real humans — is the only category requiring actual judgment.

Slack triage (/slack) — Same three-tier classification, adapted for channels and threads. Instead of email aliases, the primary signal is a channel-priority map: my team channel defaults to Tier 1, #random defaults to Tier 3, and DMs from my manager are always urgent regardless of content. What used to be a 25-minute morning scroll becomes a 30-second structured scan.

AI news digest (/news) — Scans Gmail for emails labeled “AI” and generates a skimmable morning briefing. Saves me the daily ritual of manually scanning 40 newsletters.

The morning stack is now: /email (30 seconds) → /slack (30 seconds) → /news (60 seconds) → act on Tier 1 items. Two minutes of scanning, then I’m working on the thing that matters. Not scrolling. Not skimming. Not wondering if I missed something.

And as of this week, even the manual trigger is optional. Cowork now supports scheduled tasks — cron-based automation that runs your skills on a recurring schedule. My news digest fires at 8:30 every morning before I'm at my desk. Same workflow, same output, zero keystrokes. The pattern is the same: define the workflow once, let it compound. The difference is I no longer have to remember to start it.

These two are open source (email, Slack). Clone them, swap in your aliases and channel priorities, and you have a morning triage system that compounds from day one.

The pattern across all three plugins is the same one Mollick identified: the value isn’t in the AI’s intelligence. It’s in the delegation structure — the tiers, the routing rules, the classification criteria. Management skills, applied to a markdown file.

---

Layer 5: Real Projects, Not Just Q&A

The biggest unlock was treating Cowork as a development environment, not a chat interface.

rvDNA bridge — I wanted to run my 23andMe DNA data through an open-source genomic analysis toolkit written in Rust. Problem: the toolkit expects continuous DNA sequences; 23andMe exports sparse SNP genotyping data. Different formats. In one Cowork session — about five minutes — Claude read the entire Rust codebase (4,679 lines across 13 modules), identified the data format mismatch, wrote a 500-line Python bridge, and ran the analysis. 596,007 genetic markers parsed in 1.08 seconds. On my computer. No cloud. No data leaving my machine.

Humanized ServiceNow Form — A Google Apps Script project with a full test suite and setup documentation.

2031 Strategy Comms — An internal communications package: Slack announcement, intranet post, leadership email, strategy visualization. Four different formats of the same message, tailored to each audience, produced in one session.

Continuity of Care forms — PDF form-filling for two of my kids, pulling from the family assistant skill’s medical and insurance references. The kind of paperwork that normally takes 45 minutes of phone-in-one-hand, filing-cabinet-in-the-other.

Beyond Reason newsletter — Every article I’ve published in the past month — on email triage, Slack triage, family assistant, and genomic analysis — was developed inside Cowork. Not generated. Developed. I use a voice and style guide skill that encodes my cadence, vocabulary, and anti-patterns, and Claude acts as a sparring partner: I bring the thesis and the opinions, it pressure-tests the arguments, synthesizes research, and helps me shape a rough idea into coherent narrative. The writing is mine. The thinking is mine. But the process of getting from “I have a take” to “here’s a publishable 3,000-word piece” went from days to hours.

The pattern: Claude isn’t answering questions. It’s doing multi-step work with full project context. Reading codebases, writing code, filling forms, drafting communications, organizing evidence, running analysis. That’s not a chatbot interaction. That’s what Hoffman means by “crafting your way through.” That’s what Mollick means by delegation. That’s what it looks like when a non-developer builds real things with AI.

---

The Context Architecture

If I had to distill everything into one principle: stop optimizing prompts and start building context infrastructure.

A prompt is a one-shot instruction. It evaporates after the conversation. Context infrastructure — CLAUDE.md files, skills, reference data, plugins, folder organization — persists. It compounds. Every session inherits everything you’ve built.

The stack looks like this:

1. Extensions — the nervous system (what the agent can reach)

2. CLAUDE.md — the operating manual (how the agent behaves)

3. Skills — domain knowledge (what the agent knows on-demand)

4. Plugins — workflows (what the agent does repeatedly)

5. Project files — session context (what the agent works on today)

Each layer compounds on the ones below it. My email triage plugin works better because the family assistant skill gives Claude context about who my kids’ schools are. The Slack triage works better because CLAUDE.md tells Claude which channels matter. The newsletter drafts are better because the voice guide skill defines my cadence, vocabulary, and anti-patterns.

This is the meta-pattern that Anthropic’s own data confirms: 52% of Claude users now work with it as a thinking partner, not a task-completion machine. The compound effect of structured context is what turns “use AI sometimes” into “AI is how I work.”

Translation: the ROI of an AI desktop agent isn’t in the model. It’s in the five layers of context you build around it. Get the layers right, and the model does the rest.

---

The Bigger Point

Here’s the thing nobody wants to hear: the model you’re using right now is probably good enough. The bottleneck was never intelligence. It was the absence of structured context — the preferences, the reference data, the routing rules, the institutional knowledge that turns a capable generalist into a useful collaborator. Without that, even the most powerful model is a brilliant amnesiac.

I didn’t start with 39 skills, 4 plugins, and 8 extensions. I started with one CLAUDE.md file and one skill. Every time I caught myself explaining something to Claude that it should already know, I added it to a reference file. Every time I repeated a workflow, I turned it into a plugin. Every time I noticed a gap in what the agent could reach, I installed an extension.

Two months later, my AI desktop agent knows my family, my inbox, my Slack workspace, my writing voice, my insurance policies, and my project architectures. It fills forms, triages communications, drafts articles, writes code, sends messages, controls my browser, and decodes genomes — all on my machine, all with my data staying local.

A fair question: why Cowork and not the open-source alternatives? OpenClaw still has an edge in messaging integrations — WhatsApp, Discord, native Slack — and the sheer breadth of its community-built agent ecosystem (190K+ GitHub stars worth of contributions). If you need a Swiss Army knife of integrations and don’t mind managing your own infrastructure, it’s a serious option. But for a single-user workflow where extensions, skills, plugins, and structured context compound into something that actually runs your day? Cowork is my choice — with the significant advantage of not having to worry about your gateway tokens getting exfiltrated through a compromised community package. Cowork runs in a sandboxed VM. Your data stays local. The trust model is simpler.

The tools will keep evolving. The principle won’t change: context beats intelligence, and the compound effect of structured delegation is the unlock.

Hoffman is right: we’re all becoming gamers. Each new challenge is a level, and AI is how you craft a way forward. But the game doesn’t start when you open the chat window. It starts when you build the context layer that makes the chat window obsolete.

I tell people this: if all AI development stopped tomorrow — no new models, no new features, nothing — and I had to spend $100 a month to maintain exactly this setup, I’d do it every day and twice on Sundays. Not because of what AI might become. Because of what it already gives back: ownership of my time.

The compound interest is real. But you have to make the first deposit.

---

The templates are all open source:

• Family assistant skill — modular household context

• Email triage plugin — inbox classification with alias routing

• Slack triage plugin — workspace triage with channel mapping

• rvDNA bridge — local genomic analysis from 23andMe data

Clone them. Customize them. Build your own. Stop treating your AI like a vending machine and start treating it like a workstation.

---

Go deeper with the thinkers who shaped this piece:

• Reid Hoffman, “We’re All Becoming Gamers” (LinkedIn, February 2026)

• Ethan Mollick, “Management as AI Superpower” (One Useful Thing, January 2026)

• Ethan Mollick, “A Guide to Which AI to Use in the Agentic Era” (One Useful Thing, February 2026)

• Anthropic, “Economic Index: Economic Primitives” (January 2026)

• Nate Jones, “2026 Sneak Peek: The First Job-by-Job Guide to AI Evolution” (Nate’s Newsletter)

If this was useful, share it with someone who’s still opening Claude to ask “what’s the weather.” They’re sitting on a workstation and using it as a calculator.