Tina Huang · 19:20 "The Productivity plugin is built by Anthropic themselves — probably way more sophisticated than anything you'd build yourself. Don't reinvent the wheel."
As of May 2026, Tina Huang has published the most comprehensive public tutorial on Claude Cowork. She is a former Meta data scientist who now runs Lonely Octopus, an AI education bootcamp. The video isn't just a how-to: it presents the structural thinking behind Cowork as Anthropic's B2C product and how its features connect, anchored in concrete use cases like an investment dashboard, expense analysis, and skill-encoding a brand book.
From the MEMEX editorial view, the important thing is that this outlines Anthropic's B2C strategy — its response to OpenAI Operator / Computer Use — not from official marketing material, but from the primary perspective of how a normal SMB operator or individual creator actually touches the product. Set alongside Intercom's company-wide Claude Code rollout (1,400-person org) and PFF's post-engineer org demonstration (200-person org), Anthropic's Claude platform comes into focus as a single system running through individual → SMB → enterprise.
Cowork = Anthropic's answer to OpenAI Operator, aimed at non-technical users
Tina's framing at the top clarifies the problem. "Claude Cowork belongs to a now very popular category of AI product called the local AI agent — an AI that lives on your actual computer and operates autonomously on files, apps, and tools." On its relationship to OpenAI Operator (Computer Use): "Cowork is Anthropic's interpretation. More secure, built exclusively on the Claude ecosystem, and targeted at non-technical users."
These three differentiators — (1) secure, (2) Claude-exclusive, (3) non-technical user — summarize the whole shape of Anthropic's B2C strategy. Where OpenAI takes the broad "general-purpose web agent" stance with Operator, Anthropic goes deep with "vertical integration centered on Claude." Consistent with the strategic axis MEMEX recorded in The structure of Anthropic's $350B valuation — "own model + own apps + partner integrations."
The product entry point is simple: from the same hub, you choose among Chat / Cowork / Claude Code. This is the printing-press thesis from Boris Cherny (creator of Claude Code) — "Claude Code isn't just autocomplete, it's an agent" — extended into a non-coding version.
Level 1 — file organization and auto-generated dashboards
The first thing Tina shows is a case where she asks Cowork to "organize" a desktop cluttered with 376 files. Cowork reads the contents, proposes a reorganization, and even surfaces a security warning: "oh — there's an OpenAI API key sitting on your desktop in the open." It strongly recommends deleting it or moving it somewhere secure. This is the kind of intervention only a local AI agent can perform — a web-based chatbot couldn't.
The next demo is more telling: she hands over 24 months of credit-card statements (CSV) and asks it to "analyze them, build an interactive dashboard, and give me insights and suggestions." Cowork picks the Opus model, processes 24 documents in one pass, and generates a local HTML artifact dashboard. Tina's note: "this can't be done in the Claude.ai chatbot UI — that one only lets you upload 20 documents. Cowork handles 24 months of statements, and the dashboard is local."
That single point — "exceeding the 20-document limit" — is the watershed between chatbot and local agent. It echoes Granola's "cannot one-shot it" insight about "rallying with the LLM like in tennis, in a feedback loop" — instead of aiming for a perfect one-shot, you design to hand off to a local agent and let it run.
Level 2 — the three-layer extension: Skills, Connectors, Plugins
Cowork's real value lies in its three-layer extension architecture. Tina's explanation, organized:
| Layer | Role | Examples |
|---|---|---|
| Skills | Reusable packages of instructions and knowledge (markdown folder) | Brand book apply, Canvas design, Doc authoring |
| Connectors | A connection standard for external tools (MCP-based) | Gmail, Drive, Calendar, Microsoft 365, custom MCP |
| Plugins | Combinations of skills + connectors (task-specialized) | Finance plugin (auto-generated P&L), Productivity plugin |
The thinking behind this three-layer structure follows directly from the official Skills proposal from Barry Zhang × Mahesh Murag (Anthropic) — "Skills = one folder, progressive disclosure" — landed in production form. Tina's process for building a brand-book skill matches exactly: she asks Cowork to "make this brand book into a skill," it saves it as a markdown file, and afterwards she can invoke it with "apply brand to X."
The relationship between Connectors and MCP lines up with the discussion in Nick Cooper (OpenAI) on MCP vs CLI. Tina shows Connectors' universal reach: "Gmail, Drive, Calendar, Microsoft 365 — pretty much everything. If it's missing, you can build your own with MCP." This is a consumer-product implementation of the division Supabase's Skill + MCP design drew: "MCP is connection; Skills are guidance."
The strategic meaning of plugins — why Anthropic ships them first-party
The most interesting moment in the video is the appearance of the Productivity plugin. Every time Tina starts a new project, she first installs the Productivity plugin (an Anthropic-official plugin) and runs start. "Technically it's for task management, but the reason I always install it is the memory system setup."
The Productivity plugin builds in Automatic CLAUDE.md / memory.md management A mechanism provided by Anthropic's official Productivity plugin. When the user runs /start, it auto-generates memory files inside the project (task.md, CLAUDE.md, memory.md, etc.), and the agent accumulates and references context in a sophisticated way. Per Tina Huang in her video, leaving this to the Anthropic plugin produces more advanced memory management than maintaining CLAUDE.md by hand. This plugin's existence signals that Anthropic is no longer purely a platform provider — it has begun competing as a first-party plugin developer. . As the user writes into the project, related context is automatically saved to memory. Tina's frank take: "Other tutorials might tell you to maintain CLAUDE.md or memory.md by hand. That's fine too. But this plugin is built by Anthropic — probably way more sophisticated than anything you'd build yourself. Don't reinvent the wheel."
That, from the MEMEX view, is the most important signal. Anthropic isn't simply providing a "platform" for Skills / Plugins — it's started competing as a first-party plugin developer itself. The Productivity plugin and the Finance plugin (auto-generated P&L, reconciliation, financial statements as an official package) are signs of Anthropic going directly after the "high-value use cases."
This resembles the early-App-Store dynamic between Apple's first-party apps and third-party developers. Anthropic's first-party plugins are convenient, but they may compress the third-party skill / plugin ecosystem. The skill-distribution issue Pedro (Supabase) raised — "no established registry, each vendor pushing its own plug-in format" — could collapse around an "Anthropic-ecosystem-centric" outcome as Anthropic reinforces its own first-party plugins.
Projects and Scheduled tasks — persistence and routinization
What turns Cowork from a one-off task tool into a "daily OS" is Projects and Scheduled tasks. Tina runs four projects — Lonely Octopus (her business), Content (content planning), Personal, and Investments — each operating as a persistent workspace with its own instructions + memory + connections.
A scheduled-task example: "At 7am every morning, build a brief from my calendar and email and send it to Apple Notes" — that's the entire request, and it becomes a cron task. Of Tina's running examples, the investments project is the most developed: "every morning, build a portfolio digest, extract action items against my investment philosophy, and visualize the whole thing in a mission control dashboard."
This is the consumer-side realization of the structural shift Eric Allam (Trigger.dev) framed as Durable Agents — "for 30 years, stateless compute was the core; agents are forcing us into stateful compute." Now that Cowork handles persistent state at the OS level, ordinary users get their first taste of "AI runs my daily routine."
Claude Code + Cowork combo — top-tier usage for power users
At the end of the video, Tina demonstrates combined use of Claude Code and Cowork from the same hub. "I build the workflows and dashboards in Cowork. But when the patches grow large and complex, Cowork isn't specialized as an AI coding agent. That's when I switch to Claude Code — pointing the working directory at the same project folder."
Tina's demo: she takes a dashboard she built in Cowork — "add a new tab to Mission Control so we can build a technical trading bot" — and runs the coding addition on the Claude Code side. This enables a division of labor inside one project: "non-engineers build the UI / experience in Cowork; engineers go deeper through Claude Code once complexity rises beyond a threshold."
It's the technical foundation that brings PFF's post-engineer engineering org concept — "anyone can write the spec / LDD; the agent implements" — down to the individual / SMB level. What a 200-person PFF can do, a one-person creator can now do too.
Editorial Notes — what Cowork reveals about Anthropic's B2C strategy
Three signals to read from this video.
(1) Anthropic intends to be a vertical-integration company. Where OpenAI goes "broad" with ChatGPT + Operator + API, Anthropic goes "narrow and deep" with the Claude model + Cowork (consumer) + Claude Code (developer) + first-party plugins. Consistent with the strategic axis MEMEX confirmed in the $350B valuation structure — independence that can say "no" to the Pentagon while staying in lockstep with MSFT / NVDA — a thoroughgoing "verticality on our own axis."
(2) First-party plugins mark a new phase in ecosystem competition. With the Productivity plugin and Finance plugin shipped officially, Anthropic shifts from "platform provider" to "direct competitor on the high-value use cases." For third-party skill / plugin developers this is pressure, but for users it produces the reassurance that "Anthropic itself is moving on this." The same two-sided dynamic as early App Store debates around Apple's first-party apps.
(3) Individual → SMB → enterprise on the same platform. By letting users access Cowork (non-technical individual) + Claude Code (engineer) + official plugins (work-specialized) from the same hub, Anthropic can cover everything from a "$20/month Pro user" to an "enterprise contract worth hundreds of millions a year" on one system. It connects the enterprise side — Intercom (1,400 people), Japan's three megabanks adopting Claude — with the individual-creator side (Tina) via the same go-to-market. Anthropic's intent to capture SaaS / consumer SaaS / API as three markets through one product family is becoming explicit.
Video Outline
- (00:00) What you can do with Cowork; investment dashboard intro
- (00:43) Definition of Cowork — local AI agent, relationship to OpenAI Operator
- (01:24) Installing the hub and the three tabs (Chat / Cowork / Claude Code)
- (02:09) Settings — Capabilities, Cowork, Cloud in Chrome
- (03:14) Level 1 begins — desktop file organization
- (05:25) Auto-generating a dashboard from 24 months of credit-card statements
- (08:00) Level 2 begins — skill-encoding a brand book
- (10:00) How to build skills, how to share, using others' skills
- (10:55) Connectors (Gmail / Drive / Calendar) and their relationship to MCP
- (11:42) Plugins — combinations of Skills + Connectors; official Finance plugin demo
- (13:48) Manus sponsor section (office decor)
- (14:03) Scheduled tasks — automated morning briefs
- (15:09) Level 3 — Projects (persistent workspaces)
- (17:01) Memory auto-management via the Productivity plugin
- (19:08) Investment project demo, the mission control dashboard concept
- (20:33) Claude Code + Cowork combo
- (22:00) Close