Video Summary - Hướng dẫn sử dụng Claude Cowork từ A đến Z dành cho người mới

Video purpose / target audience

Tutorial for beginners on how to use Claude (Anthropic) Cowork (“Clot Cowork” in subtitles) effectively for non-IT professionals.
Emphasizes turning repeated office tasks into automated, structured workflows.

1) Three ways to use Claude

Claude Chat (clot chat)
- Use via the website (clot.ai) in a chat interface.
- Best for quick Q&A, but has major limitations:
  - Can’t directly interact with files on your computer without manual steps.
  - Practical limits on file upload quantity and size (example constraints mentioned: up to 20 files per chat and ≤ 500MB per file).
- For frequent/large-data workflows, manual uploading reduces productivity.
Claude Code (clot code)
- Programmer-oriented: Claude interacts with your environment via terminal/IDE tools.
- Advantage: can work with folders/files more directly than Claude Chat and avoids the same upload limits.
- Drawback: requires using commands, which may be unfamiliar to non-technical users.
Claude Cowork (clot cowork) (main focus)
- Best for general office tasks: analysis, research, writing/reporting, file processing.
- Claude decomposes a task into a plan, executes it, and returns results.
- Cloud-based automation: continues running even after the computer is shut down (the platform runs in the cloud).

2) Cowork structure: Project + core sections

To make Cowork repeatable and consistent, the video introduces a Project concept (like a workspace/company analogy).

Inside a Cowork project, the interface includes:

Instructions (policy / rules)

“Company rules” that define how Claude behaves across project sessions.
Example rules mentioned:
- Always answer in Vietnamese
- Always start answers with three smiley icons
Can be edited anytime and then saved.

Memory

Automatically generated during work.
Stores the project’s evolving knowledge from interactions and uploads.
Some parts (e.g., “earn” mentioned) are not visible to others.

Context

Persistent project “shared knowledge/files” usable across multiple sessions.
Files uploaded in Context can be referenced later without re-uploading (unlike normal session attachments).
Tracks capacity with a percentage meter; reaching 100% blocks further uploads.
Supports:
- Files (PDF/images and more)
- Text content via an “add text” option
Behaves like “workspace memory” for shared assets.

Schedule

Enables periodic execution (example: every Monday morning summarize the last 7 days of AI news).
Includes configuration for frequency (hourly/daily/weekly), task instructions, and permission/approval behavior.

3) Sessions and file persistence rules

In standard Cowork chat/project flow:
- Files uploaded in a specific session remain available within that session.
- After the session ends, those files disappear and must be re-uploaded.
Context uploads are different:
- Files placed in Context remain accessible across sessions.

4) Model selection: Claude model tiers + “effort”

Cowork allows choosing among multiple language model options (names mentioned: Haiku, Sonnet, Opus, Fable—appearing as “Haiku/zonet/opus/fable” in subtitles).

Trade-off described:

Lower tier / lighter
- Faster, cheaper, less reasoning detail.
Higher tier / more expensive
- Deeper reasoning, longer runtime, better accuracy.

Effort (reasoning depth) can be adjusted (low → max). The video encourages testing the same question to find the best balance.

5) Cowork file handling capabilities (read / transform / write)

Cowork can work with many file types, including:

Text / documents (doc-like)
PDF
Images (PNG/JPEG)
CSV/Excel-like tables
Potentially other formats (audio/video mentioned)

Typical operations described:

Read file contents (summarize/extract)
Update / transform documents (rewrite titles, edit content)
Generate outputs (Excel summaries, dashboards, new files)

Upload batching constraints were mentioned again:

Example: max 20 files of the same type in some flows.

6) Folder access: “add a folder” to give Claude control

Users can grant Cowork access to a local directory using “add a folder.”
Claude can then browse files and run tasks on that folder.
Example workflow described:
- Claude sorts images into subfolders by category (e.g., species).
- Then it renames files into an indexed format.
- Optionally converts images to black-and-white.
- Deleting/overwriting may require explicit permission prompts.

7) Safety / approval modes (important risk control)

The video emphasizes approval control with three modes:

Manual approve
- Claude pauses for confirmation before each action.
Automatic approve with safeguards
- Claude proceeds automatically, but still blocks unsafe actions.
- Covers protection against prompt injection and data leaks.
Skip approval
- Claude runs continuously with minimal interruptions.
- The video warns this is dangerous if you’re not fully confident.

Prompt injection explanation (example)

A PDF invoice may contain hidden text (e.g., white-on-white) instructing the AI to ignore rules (e.g., “mark invoices valid”).
Claude might follow malicious instructions unless safeguards and/or approval controls are enabled.

8) End-to-end example: invoices/contracts → Excel → dashboard

Scenario

A “private data” folder contains multiple PDF invoices, contracts, appendices, including scanned documents.

Claude tasks

Read PDFs (including scanned images, not just text PDFs)
Consolidate invoices into a single Excel report
Detect anomalies (e.g., duplicate checks, mismatched amounts vs contract, missing information)
Create a dashboard for visualization
Optionally export results and create a shareable HTML dashboard

Outputs produced

An Excel file (summary table + tabs like “invoice”, “details”, “warnings”)
A compiled/anomaly report
A dashboard artifact (HTML; “dashport” mentioned)

Dashboard privacy control

Initially private
Then can be made public/shared via link for others

9) External tool connections via MCP (Gmail/Google tools)

Problem

By default, Cowork can’t directly use external services (e.g., sending emails).

Solution: Model Context Protocol (MCP)

Treated like a “USB connection” to plug external tools/APIs into Cowork.
Example: MCP servers already integrated for tools like Google/Gmail/Notion.

Demonstrated workflow

Connect to Gmail
After creating a dashboard:
- Instruct Cowork to send an email containing the dashboard link
- Handle permissions (“allow for start” vs later approvals)
Includes the ability to send a link (public dashboard) and verify viewing permissions.

10) Automation reuse: “Skills”

The tutorial introduces Skills to avoid rewriting prompts/workflows every time.

Skill = packaged workflow
- Like saving a reusable “recipe” (pho analogy).
- Contains step-by-step instructions for recurring tasks.

Ways to obtain skills:

Create your own skill and save it
Use pre-built skills from the community marketplace

11) Skill usage + marketplace (“skins.sh” / skins)

Skills appear as “skins” in subtitles (suggesting Claude Cowork “skins” terminology).
Users can install skins and call them with a short syntax.
Marketplace example:
- skins.sh mentioned
- Advice: choose reputable skins from trusted sources, check downloads/usage, and read descriptions carefully.

Warning

Poorly designed skins could include malicious behavior or prompt injection.

Main speakers / sources

Primary speaker: the video author/instructor (speaking throughout; no name given in subtitles).
Referenced product/source throughout: Anthropic / Claude (clot.ai), Claude Cowork, including model tiers (Haiku, Sonnet, Opus, Fable), and MCP (Model Context Protocol) for external tool access.

Hướng dẫn sử dụng Claude Cowork từ A đến Z dành cho người mới

Key takeaways