Cowork Fundamentals in 22 Minutes — Tina Huang's Level 1 → Code Walkthrough

Source: Claude_Cowork_Fundamentals_In_22_Minutes Creator: Tina Huang URL: https://www.youtube.com/watch?v=uGwDuvSqgYI Duration: ~22 minutes Platform: YouTube

A 22-minute beginner-to-advanced Cowork walkthrough by Tina Huang (Lonely Octopus founder) that escalates from “organize my messy desktop” to “self-running investment mission-control dashboard” across five capability levels — file organization, skill creation, connectors/plugins, scheduled tasks, and Projects with the official Productivity Plugin as a memory primitive. The pedagogical spine is the level escalation itself; the load-bearing recipe is the Productivity Plugin → don’t reinvent your own CLAUDE.md / memory.md pattern. Closes with Cowork + Claude Code as the “top 0.1% of users” combination.

Key Takeaways

• Cowork = Anthropic’s interpretation of OpenClaw. Positioned in the “local AI agents” category — an AI that lives on your actual computer and autonomously acts on files, apps, and tools. Pitched explicitly as “more secure, built using the Claude ecosystem, targeted at non-technical users” vs OpenClaw.
• Settings to enable on day one. Capabilities tab: artifacts, AI-powered artifacts, inline visualizations all ON. Cowork tab: Dispatch ON (phone control). Claude in Chrome ON. Global instructions field: insert a personal context block (“who I am, how to work with me, how to build, what to never do”) that applies to every Cowork session — there’s a meta-prompt to generate this for you and paste back into settings.
• Level 1 — file organization. Point Cowork at a messy local folder, ask “help me organize these files.” Sonnet 4.5 default. Cowork proposes a folder structure as a question, surfaces concerns (e.g. “I noticed an exposed openai.txt API key — recommend deleting it”), executes after approval. Differentiator from claude.ai chat: chat can’t modify files on the actual computer.
• Level 2 — credit-card statement analysis. Drop 24 months of statement CSVs into a folder; ask Cowork (Opus, “lowkey important”) to build an interactive spending dashboard. Opus surfaces three clarifying questions (output format, business-vs-personal treatment, insight priorities) before executing. Output is a local HTML artifact with monthly breakdown, top merchants, fees-and-interest flags, subscription tracking. Why Opus matters here: the chat-only path is capped at 20 documents uploaded — Cowork has no document cap because it reads from the folder.
• Level 2 (cont.) — skill creation by example. After building a brand book from local brand-asset screenshots (Lonely Octopus logo + website screenshots → PDF brand book via “make this comprehensive”), Tina says: “make applying the brand book onto visual things we build a skill.” Cowork generates a reusable Skill (text file with the brand-application instructions) and stores it locally. Invoke later by typing “apply brand to [thing].” Skill scope can be local or shareable.
• Connectors and plugins distinct from skills. Connectors = third-party integrations (Gmail, Drive, Calendar, Microsoft 365, MCP-based custom ones). Plugins = bundled skills + connectors for specific verticals — example: the Anthropic-built Finance plugin ships audit-support / close-management / financial-statement skills plus Snowflake / Databricks / BigQuery / Slack / Gmail connectors. Activated by typing /finance then picking a skill. Anthropic-built Finance plugin highlighted as “game changer” for business operators.
• Scheduled tab — set-and-forget routines. schedule every morning at 7am, give me a brief of my day based on my calendar and a short brief of any emails, send to Apple Notes. Activated by going to Scheduled tab → click task → Run. Tina uses Apple Notes specifically so the brief is on her phone when she wakes. This is the “intermediate user” promotion gate.
• Level 4 — Projects as persistent workspaces. A Cowork Project = folder + project instructions + memory across sessions. Tina’s setup: parent folder Cowork Playground with sub-projects Lonely Octopus (curriculum/bootcamp tracking), Content (content idea research), Personal (personal problems), Investments (investment workflows). Each project’s folder mirrors the Cowork Project view 1:1.
• The Productivity Plugin = memory-system scaffold. Tina’s load-bearing recipe: “Whenever I start a project, I install the Productivity Plugin and run /productivity start.” Even when not managing tasks per se, the plugin sets up the memory system — creates task.md, dashboard, memory directory, and a CLAUDE.md / memory.md scaffold. The explicit reasoning: “you may have seen tutorials telling you to maintain CLAUDE.md or memory.md yourself. You could. But this plugin does that for you and is more sophisticated than what you’d figure out — Anthropic built this. Don’t reinvent the wheel.” After the plugin is installed, dropping a screenshot of stock positions and saying “here are my investments” → Cowork writes everything to memory automatically.
• /update keeps memory current. After the Productivity Plugin is initialized, typing /update (in any future session of that project) prompts Cowork to fill in any gaps in the memory it has — refreshes the agent’s working knowledge of the project state.
• Mission Control dashboards as the Project payoff. With Project memory + connectors + skills + scheduled tasks composed, a single “Mission Control” HTML dashboard renders: investments, daily portfolio digest, new opportunity research, alignment checks against personal investment philosophy. Each project gets its own dashboard.
• Cowork + Claude Code = the 0.1% combination. Closing pitch: once Cowork-built projects get bigger and more complex, switch to Code tab in the same Claude Desktop hub, change the working directory to the project folder, and start extending the Mission Control dashboard via Claude Code. Tina’s example: “add to my mission control dashboard a new tab that lets me create technical trading bots.” The handoff is just the tab switch — same project folder.

The Five-Level Escalation Framework

Level	Capability	Trigger	Example
1	File operations	Point at folder, plain-English ask	”Organize my desktop”; “Analyze my credit card CSVs and build a dashboard”
2	Skills (reusable instructions)	“Make this a skill” after Cowork does it once	”Apply brand to [thing]” — extracted from one-off brand book build
3	Connectors + Plugins	Activate via + → Plugins / Connectors	Finance plugin → /financial-statements generate Oct 2025 P&L
4	Scheduled tasks	schedule [task] every [cadence]	7am daily brief from calendar + emails → Apple Notes
5	Projects + Productivity Plugin	New project → /productivity start → memory scaffold	Investment mission-control with daily digest + opportunity research
5+	Cowork → Claude Code handoff	Switch to Code tab in same hub, same folder	Extend Mission Control dashboard with trading-bot tab via Claude Code

Implementation

• Tool/Service: Claude Cowork (claude.com/product/cowork) in Claude Desktop.
• Setup:
  1. Download Claude Desktop, install. Three tabs appear: Chat / Cowork / Code.
  2. Settings → Capabilities: enable artifacts + AI-powered artifacts + inline visualizations.
  3. Settings → Cowork tab: enable Dispatch (phone control) + write Global Instructions block (use a meta-prompt to generate).
  4. Enable Claude in Chrome extension.
  5. First task: point at messy folder, ask “help me organize these files,” Sonnet 4.5.
  6. After two successful operations, capture one as a Skill (make [pattern] a skill).
  7. Browse Plugins (+ → Plugins) — install Productivity, Finance, etc. as needed.
  8. New Project: parent folder + sub-folder + project instructions (“You are my financial analyst. Push back on me. Be direct.“) + /productivity start for memory scaffold.
• Cost: Cowork included in Pro/Max/Team/Enterprise. Tina uses Sonnet for simple tasks, Opus for high-stakes analysis (statement analysis, multi-step research).
• Integration notes: Desktop and computer must stay awake during Cowork execution. Dispatch lets you trigger from phone but execution still happens on desktop. The Productivity Plugin is the highest-leverage Anthropic-built plugin for personal AIOS users.

The Productivity Plugin as Memory Primitive

Tina’s strongest contribution to the Cowork-as-AIOS pattern: she explicitly recommends stop building your own CLAUDE.md / memory.md scaffolds for Projects — install the Productivity Plugin and let it manage memory for you.

Without Productivity Plugin	With Productivity Plugin
Manually maintain CLAUDE.md + memory.md per project	Plugin creates and maintains them via /productivity start
Bespoke memory schema per project	Sophisticated default schema built by Anthropic
Manual updates on context shift	/update auto-fills gaps
Vulnerable to schema drift across projects	Consistent across all projects
Higher cognitive load to maintain	”Don’t reinvent the wheel — Anthropic built this and they’re better at it than you”

This is a notable counter to the Jeff’s Jarvis build pattern (manual three-level CLAUDE.md hierarchy) and the Ben’s five-skill AIOS pattern (custom /os-setup). Tina’s Operator-track recommendation: don’t custom-engineer the memory layer — use the Anthropic plugin.

Try It

1. First Cowork session today: install Claude Desktop, enable artifacts + Dispatch + Claude in Chrome, run “help me organize the screenshots folder” with Sonnet.
2. Capture your first skill: after Cowork does any branded task once (e.g., write copy in your voice), say “make this a skill” — verify it appears in Customize → Skills → Local.
3. Install Productivity Plugin before your next Project: + → Plugins → Productivity → install. Run /productivity start. Note the task.md + memory directory it creates.
4. First scheduled task: schedule every morning at 7am, give me a brief of my day based on my calendar and a short brief of any emails, send to Apple Notes. Activate from Scheduled tab.
5. First Cowork + Claude Code session: open an existing Cowork Project, switch to Code tab in the same hub (don’t switch to a different app), keep the same working directory, ask Claude Code to extend something you built in Cowork.

Related

• Claude Cowork (Product Overview) — the canonical product article; Tina’s video is the operator-track companion walkthrough
• Getting Started with Cowork (Notion Walkthrough) — the Eliot Prince 7-step + 5-use-case sibling walkthrough; complementary because that one is depth-on-use-cases, this one is depth-on-escalation
• Cowork Jarvis Build (Jeff) — three-level CLAUDE.md hierarchy alternative to the Productivity Plugin
• Ben’s Five-Skill AIOS Setup — /os-setup custom skill alternative to the Productivity Plugin
• Cowork “AI Consultant” Recipe — Eliot Prince’s 4-knowledge-file Project pattern for client work
• Seven Cowork Live Artifacts Dashboards (Eliot Prince) — Mission Control dashboard variants for daily ops surfaces
• Cowork Plugins — Finance plugin and the broader plugin marketplace
• Cowork Dispatch — phone-to-desktop trigger pattern Tina enables in her settings setup

Open Questions

• Productivity Plugin scope. Is the plugin truly Anthropic-built and -maintained, or third-party? Source verbally attributes it to Anthropic but does not show the publisher tag.
• /update mechanism. Does /update re-scan the project folder for newly added files, or only update the memory schema based on the active session’s new context?
• Dispatch when desktop sleeps. Tina enables Dispatch but doesn’t demo the “phone command while laptop is asleep” case — does Dispatch still hold the task until the desktop wakes?
• Multi-project Productivity Plugin state. Each project gets its own Productivity Plugin install (per the recipe), or is there cross-project memory sharing in any form?