You've got the basics. Now learn to build skills, create plugins, automate with Chrome, orchestrate agents, and turn Claude into your operating system.
Message box shows a Chat / Cowork switch? Pick Cowork for this. Why.
If you followed the Level 1 guide, you've already been using skills without knowing it. When you say "new project" and Claude creates folders, memory files, and output directories, that's a skill in action. The "new project" workflow, the file naming conventions, the per-project memory, all of these are patterns that could be (and are) encoded as skills. Now you'll learn to build your own.
A skill is a markdown file called SKILL.md that teaches Claude a specific capability. Every time you find yourself giving Claude the same instructions, "format reports this way," "research companies using this framework," "always check these 5 things", that's a skill waiting to be built.
You give the same instructions 3+ times. You have a specific workflow or framework. You want consistent output quality across sessions. You want to share a process with colleagues.
It's a one-time task. The instructions are simple enough to type each time. The process changes frequently and isn't standardized yet.
Every skill has two parts: YAML frontmatter (metadata) and markdown body (instructions).
description field. Be specific about trigger phrases. Vague descriptions = skills that never fire. Include examples of what users might say: "research [company]", "tell me about [company]", "competitive profile".
| Location | Scope | Priority |
|---|---|---|
~/.claude/skills/ | All your projects (personal) | High |
.claude/skills/ in project | That project only | Medium |
| Inside a plugin | Anyone who installs the plugin | Standard |
| Enterprise managed | Entire organization | Highest |
Type /skill-creator in any session, or ask Claude to help you build a skill. It is a skill from Anthropic. Claude will interview you about what you want the skill to do, write the SKILL.md for you, test it, and optimize the description for reliable triggering. You don't need to write YAML by hand.
!`git log --oneline -5` injects your recent commits. !`cat package.json | jq .dependencies` injects your project dependencies. This happens before Claude sees the skill, so the data is always fresh.
A plugin is a folder that bundles skills, commands, and connector configurations into a single installable package. If skills are individual tools, plugins are the toolbox.
plugin.json goes inside the .claude-plugin/ directory. Skills, commands, and agents go at the plugin root level. This is the most common mistake when building plugins.
Push your plugin folder to a GitHub repo. Anyone installs with: /plugin marketplace add your-username/your-repo. Free, versioned, and easy to update.
Zip the plugin folder as a .plugin file and email or share it. Recipients add it under Customize → Plugins → Add → Upload plugin.
Submit the repo through Anthropic's directory submission portal, open to paid plans since September 25, 2026. Each submission is validated and safety-scanned, then reviewed. Once approved, you choose when to publish, and you get install and search analytics.
On the receiving end of a plugin? Claude Plugins has the check to run before you install one.
On August 12, 2026, Claude in Chrome stopped being a separate extension. The side panel is now a full Claude session, the same one you run on desktop, web, and mobile. Your skills and connectors work inside it, which they did not before. Start a task in Chrome, finish it on your desktop, the context carries over. Claude still sees, clicks, types, and navigates any page. It does that as the same Claude, not as a bolt-on tool.
Since August 26, 2026 the desktop app also has its own browser. It opens in a side panel when a task needs a website. It is Claude's browser, separate from yours, and it starts with none of your logins. You bring them over site by site, and the import leaves out banking, email, and single sign-on sites unless you include them.
Pick by where the work lives. Use the built-in browser when you want Claude working on the web while you keep using your own Chrome. Use Claude in Chrome when the job needs the pages and accounts you already have open. If you already use Claude in Chrome it stays the default, and you can change the preferred browser in the desktop app's settings.
Claude manages multiple tabs simultaneously. Drag tabs into Claude's group, and it can read and act across all of them. "Compare pricing across these 5 competitor websites."
Claude fills forms using context it already has. "Fill out this vendor registration form using our company details." Works with any web form.
"Go to LinkedIn and pull the last 20 posts from these 5 people into a spreadsheet." Claude navigates, reads, extracts, and organizes.
Record browser interactions as GIF animations with click indicators and action labels. Perfect for creating demos, tutorials, and documentation.
Claude reads console logs, DOM state, and network requests directly. "Check why the login page is throwing a 403", it reads the actual error.
Combine Claude in Chrome with scheduled tasks. "Every morning at 9am, check these 3 dashboards and summarize what changed."
Claude reaches a service through a connector when that service has one. It drives the page through Chrome when the service does not. But what about everything else, Finder, Notes, Maps, System Settings, Photoshop, Excel (the desktop app), or any native application? That's Computer Use. Claude takes screenshots of your screen, moves the mouse, clicks buttons, and types, just like a remote coworker sitting at your desk.
This is the mental model that makes everything click. When you give Claude a task, it picks the fastest, most reliable tool available:
Connectors are fast and precise, they use each service's API. The browser, yours through Claude in Chrome or Claude's own built-in one, handles web apps without a connector. Screen control is the fallback for everything else.
Computer Use is off by default. To turn it on:
Claude always asks permission before touching a new app. You approve each application explicitly, and you can stop it at any time.
"Organize my Downloads folder, move PDFs to Documents, images to Photos, delete anything older than 30 days." Claude opens Finder and does it.
"Pull the metrics from this dashboard in Chrome, drop them into the Excel spreadsheet on my desktop, then email the updated file." Three apps, one prompt.
"Fill out this insurance application in Preview using my info from the About Me file." Claude reads your context and types into native form fields.
"Check my disk space, clear the trash, and tell me what's taking up the most storage." Claude navigates System Settings and Finder to investigate.
Computer Use is live, you can redirect Claude while it works. These steering prompts are your remote control:
| Steering Prompt | What It Does |
|---|---|
"Stop. Don't click that yet." | Pauses before a risky action |
"Use Chrome, not Safari." | Redirects to a different app |
"Take a screenshot and show me before you send it." | Review checkpoint before proceeding |
"Skip Slack. Open the browser instead." | Overrides the default tool choice |
"Actually, start over. New approach." | Full reset mid-task |
Combine Computer Use with Dispatch for fire-and-forget workflows:
This works for any Computer Use task. Export a deck to PDF, attach it to a calendar invite, organize files, all while you're at lunch.
Claude can run tasks on a schedule, daily, weekly, or at specific times. Combined with plugins and Chrome, this is where Claude becomes an autonomous operator.
| Schedule | Cron Expression | Plain English |
|---|---|---|
| Every morning | 0 9 * * * | 9:00 AM daily |
| Weekdays only | 0 9 * * 1-5 | 9:00 AM Mon-Fri |
| Every Monday | 30 8 * * 1 | 8:30 AM Mondays |
| First of month | 0 0 1 * * | Midnight, 1st day |
Daily at 9am: summarize unread emails, check calendar, create a priorities list with talking points for today's meetings. Outputs a markdown file you review with coffee.
Fridays at 5pm: scan email for invoices received this week, extract amounts and vendors, add to a tracking spreadsheet, flag anything over $5K for review.
Every Monday: check 5 competitor websites for pricing changes, new features, or blog posts. Compile a diff report showing what changed since last week.
Claude's memory isn't one thing, it's a hierarchy. With the March 2026 Projects update, there are now four distinct layers. Understanding them lets you put the right information in the right place.
Where: Claude Desktop → Settings → Personal Preferences
Scope: Every conversation, every project, every session
What goes here: Your name, role, communication style, personality preferences, behavioral rules. Things that are always true about how you want Claude to work.
Where: Project → Instructions field + automatic memory.md
Scope: All tasks within that project
What goes here: Project-level tone, formatting rules, and key context in the Instructions field. Auto-memory captures decisions and patterns as you work, Claude decides what's worth saving. First 200 lines (or 25KB) load at the start of each new task.
Where: CLAUDE.md file in your workspace folder (or subfolder)
Scope: That specific folder only
What goes here: Detailed folder protocols, naming conventions, session start behavior, project registry, read-only/write-only rules. More structured and opinionated than native Instructions, this is your operating system.
Where: MEMORY.md in your workspace folder
Scope: Persistent across sessions, user-controlled
What goes here: Decisions, deadlines, status updates, things you specifically want remembered with your exact wording. Say "remember this" to write; "forget that" to delete. Unlike auto-memory, nothing gets saved unless you ask.
This is the most common question: "If Projects auto-remember things, why do I need MEMORY.md?" Because they serve different purposes:
| Native Auto-Memory | MEMORY.md (your workspace) | |
|---|---|---|
| Trigger | Automatic, Claude decides | User-triggered, "remember this" |
| Content | Whatever Claude thinks is useful | Exactly what you told it to save |
| Structure | Sorted into Topics you can open in Settings → Memory | Your format, your categories |
| Control | Medium, Claude curates and you can edit or delete a Topic | High, you curate |
| Risk | May save noise or miss critical facts | You might forget to say "remember this" |
| Best for | Background context, patterns, preferences | Key decisions, deadlines, exact specifications |
Both native Projects and a workspace folder support project-scoped isolation. Each native Project has its own auto-memory. Each workspace subfolder can have its own CLAUDE.md and MEMORY.md. Your marketing project's Claude has different context than your finance project's Claude.
@include line so the main file stays a short table of contents. The deeper habit is separating what from how: keep your reusable instructions in files, keep one-off asks in the chat. When the two blur, both get harder to fix.
The setup prompt in the Beginner guide scaffolds a version of this in one conversation, and most people should let it. This section is for when you want to understand exactly what it builds, or set it up manually and tune every piece. The memory layer that sits on top, CLAUDE.md and MEMORY.md, is covered in Memory Architecture Mastery above; this is the folders and the identity file underneath it.
The core idea is one separation: where Claude reads from versus where Claude writes to. Native Projects don't enforce this, so without a structure your inputs and outputs pile into one folder. Here's the layout power users run, all of it inside the cloud-synced drive you picked during setup:
Your identity file. Claude reads it at the start of every session. This is what makes Claude feel like your assistant instead of a generic AI. Detailed below.
Document structures Claude should follow: a report format, an email style, a deck skeleton. Claude studies the pattern and applies it without copying the content. Start empty; add as you find patterns you reuse.
Context Claude needs: briefs, research, past work, organized by project. Claude reads from here but never writes to it, so your originals stay untouched.
Every file Claude creates lands here, mirroring your PROJECTS folders. Keeps the workspace clean and makes deliverables easy to find.
PROJECTS/[name]/ and a matching CLAUDE OUTPUTS/[name]/ with its own project-scoped memory.
Use one pattern for everything Claude creates: Project_ContentType_v1.ext
| Example | What It Is |
|---|---|
Acme_Proposal_v1.docx | First draft of a proposal for Acme |
Q2-Launch_Deck_v2.pptx | Second revision of the Q2 launch deck |
Monthly_Report_v3.xlsx | Third revision of a monthly report |
Ask Claude to revise something and it increments the version (v1 → v2 → v3), giving you a built-in history with no git required.
.md files. Double-click one and it opens in TextEdit as a mess of hashtags and asterisks, so most people quietly stop editing their context, and their setup never gets smarter. Fix it with Obsidian: free for personal use, runs locally, renders markdown cleanly. Download it, choose "Open folder as vault", point it at your workspace folder. Now you browse and edit context in Obsidian and generate new work in Claude, same files, no syncing. Not Notion (it corrupts markdown with hidden formatting) and not Google Docs (constant conversion). Obsidian just reads your folder without touching it.
CLAUDE OUTPUTS/ weekly, or Claude starts reading stale drafts as context.
A plain text file in ABOUT ME/ that tells Claude who you are and how you work. Without it, every session starts cold. With it, Claude works like a colleague who's known you for months. The structure:
If you've used ChatGPT for months, you've already built up your role, preferences, and corrections. Extract it instead of starting from scratch.
[email protected] with a link that expires in 24 hours. Check spam.
Unzip the export (you'll get chat.html), drop it in your workspace, and tell Claude:
"I just exported my ChatGPT history. It's in chat.html. Read through it and extract: (1) everything it knew about me, role, company, preferences, (2) instructions I repeated often, (3) recurring projects, (4) key decisions and facts. Use what you find to draft my About Me and seed my MEMORY.md."
What took months to teach ChatGPT, Claude picks up in one conversation.
MCP (Model Context Protocol) is how Claude connects to external services. Every connector you use, Google Drive, Slack, Gmail, GitHub, is an MCP server under the hood. You can build your own.
Your company's CRM, ticketing system, internal wiki, or custom database. If it has an API, Claude can connect to it.
"Pull this week's support tickets from Zendesk, categorize them, and create a summary for the team standup." Requires a Zendesk MCP connector.
| Type | Where It Runs | Best For |
|---|---|---|
| Desktop | Your local machine | Local databases, file systems, local tools |
| Remote | Cloud server (HTTPS) | Web APIs, shared team tools, multi-device access |
When Claude encounters a task that can be broken into independent pieces, it automatically spawns sub-agents, parallel workers that execute simultaneously. This is one of the most powerful features most people never discover.
Say you ask: "Research these 8 companies and create a competitive matrix." Instead of researching them one by one (slow), Claude spawns 8 parallel agents that each research one company simultaneously. Results come back in minutes instead of an hour.
Include this in your request when the task involves multiple independent workstreams. Claude will automatically identify what can run in parallel and split the work. You can also include this in your Personal Preferences or CLAUDE.md so Claude always looks for parallelization opportunities.
"Research these 5 companies", each company gets its own agent researching simultaneously.
"Analyze these 10 spreadsheets and summarize findings", each file gets its own agent.
"Write social posts for each of these 8 topics", parallel content generation.
"Update the frontend, backend, and tests for this feature", three agents working different parts.
| Type | What It Does | Best For |
|---|---|---|
| general-purpose | Full capabilities: search, read, write, execute | Most tasks |
| Explore | Read-only: fast search and analysis | Research, codebase exploration |
| Plan | Decision-focused: architecture and strategy | Planning before execution |
| File Type | What You Get | Example Prompt |
|---|---|---|
| Excel (.xlsx) | Formulas, formatting, charts, multiple sheets | "Create a financial model with revenue projections" |
| Word (.docx) | Headings, TOC, page numbers, letterhead, tables | "Write a 10-page market analysis report" |
| PowerPoint (.pptx) | Slides, layouts, speaker notes | "Build a 15-slide investor pitch deck" |
| Formatted documents, forms, merged files | "Merge these 3 PDFs and add page numbers" | |
| HTML | Interactive dashboards, reports, web tools | "Create a dashboard from this CSV data" |
| CSV/Data | Cleaned data, analysis results, exports | "Clean this messy CSV and export with headers" |
Upload a CSV and ask: "Create an interactive HTML dashboard with filters, charts, and a summary table." You get a single-file HTML that works in any browser, no server needed. Share it with anyone. Since October 8, 2026, paid plans also have Claude Dashboards, which builds it inside Claude and shows the query behind every number; both routes are in Workflow 5.
Upload a screenshot of a design mockup and ask: "Convert this to a working React component." Claude's vision analyzes the layout, colors, and spacing and produces functional code.
Drop 5 research documents into your workspace and ask: "Read all of these and create a one-page executive summary highlighting the key insights and contradictions across them."
Claude can now create interactive charts, diagrams, and visualizations inline in its responses. Ask "Show me a chart of our monthly revenue" and get a live, clickable chart right in the conversation, not just a static image. These can be exported as standalone HTML files you share with anyone.
Claude now lives inside Excel and PowerPoint as native add-ins. Ask questions about your spreadsheet, generate formulas, build pivot tables, or create slides, without leaving the app. Context carries between both add-ins, so Claude understands your data when you switch to building a deck from it. This is a third way to use Claude alongside the desktop app and browser extension.
Each plugin category transforms Claude from a generalist into a domain specialist. Here's what the major categories enable:
Account research that pulls web data + CRM context into a one-page brief. Call prep that builds agendas with attendee research. Competitive battlecards as interactive HTML with comparison matrices. Pipeline review that flags at-risk deals and creates weekly action plans. Outreach drafting with personalization from research data.
Journal entries with proper debits/credits and supporting detail. Account reconciliation matching GL to subledger/bank. Variance analysis decomposing budget-vs-actual into drivers. Financial statements with GAAP presentation and period-over-period comparison. SOX compliance testing workpapers and sample selection.
Contract review against your negotiation playbook with clause-by-clause analysis. NDA triage classifying as GREEN/YELLOW/RED in seconds. Compliance tracking for GDPR, CCPA, and data subject requests. Risk assessment with severity-by-likelihood framework.
Campaign planning with objectives, channel strategy, and content calendars. SEO audits covering keyword research, on-page analysis, and competitor comparison. Email sequences for nurture flows and onboarding. Brand voice review checking content against your style guide. Performance reports with KPIs and optimization recommendations.
PRDs with problem statements, user stories, and acceptance criteria. Roadmap management using RICE, MoSCoW, or ICE frameworks. User research synthesis turning interview notes into themes and opportunity areas. Stakeholder updates tailored to audience (exec, engineering, customers). Competitive analysis with feature comparison matrices.
SQL queries optimized for your warehouse dialect (Snowflake, BigQuery, Postgres). Data exploration profiling datasets for quality, distributions, and outliers. Visualization with matplotlib, seaborn, or plotly. Interactive dashboards as self-contained HTML with Chart.js filters. Statistical analysis including hypothesis testing and trend detection.
Create a private GitHub repo as your org's plugin marketplace. Team members install from it directly. Control which plugins are available, which are auto-installed, and which are hidden.
Admins can auto-install plugins across all team members. New hires get your org's skills on day one without any setup. Managed plugins take highest priority, they override personal settings.
Track usage, costs, and tool activity per team via OpenTelemetry integration. See which skills get used most, which teams are heaviest users, and where to invest in more automation.
Customize the home screen with your company logo and branded experience.
Dispatch is a persistent thread between your phone and your desktop Claude. Send a task from anywhere, the airport, a meeting, your kid's soccer game, and Claude executes it on your Mac using your files, connectors, and plugins. It's not cloud-based: your files never leave your computer.
1. Open Claude Desktop app on your Mac
2. Click Dispatch in the left sidebar
3. Connect your phone (iOS or Android) via the Claude mobile app
4. Send a task from anywhere. Come back to finished work.
Unlike starting a new session, Dispatch maintains a single continuous thread. Claude retains context from your previous tasks, no re-explaining.
Since March 2026, Dispatch includes Computer Use. Claude can open apps, move files, fill forms, and navigate your desktop while you're away.
"Export the pitch deck as PDF and attach it to my 2pm meeting invite." Send it, pocket your phone. Claude opens the app, exports, attaches. Done.
Queue up 5 tasks before breakfast. Review the results over coffee. Entire morning routines, email summaries, calendar prep, report generation, handled before you sit down.
The Claude mobile app now supports fully interactive applications, live charts, sketch diagrams, and shareable visual assets rendered directly in mobile conversations. It also integrates with work tools like Figma, Canva, and Amplitude from your phone. Combined with Dispatch, this means you can review designs, build interactive dashboards, approve assets, and trigger desktop work from anywhere.
| Repository | What It Has |
|---|---|
| anthropics/knowledge-work-plugins | Official Anthropic plugins: Sales, Finance, Legal, Marketing, Product, Data, Design, Operations |
| anthropics/skills | Official skills reference and examples |
| travisvn/awesome-claude-skills | Community-curated list of 37+ skills |
| ykdojo/claude-code-tips | 45 tips for getting the most out of Claude Code |
| Topic | Link |
|---|---|
| Skills Documentation | docs.claude.com/skills |
| Plugin Development | docs.claude.com/plugins |
| Chrome Extension | code.claude.com/chrome |
| Scheduled Tasks | code.claude.com/scheduled-tasks |
| Memory System | code.claude.com/memory |
| Hooks & Automation | code.claude.com/hooks |
| Agent Teams | code.claude.com/agent-teams |
| MCP Connectors | modelcontextprotocol.io |
Build your first skill. Share it with a friend. Automate something you do every week. The difference between "using Claude" and "operating with Claude" is everything on this page.