클로드 채팅·코워크·클로드 코드 완벽 정리 | 프로젝트·커넥터·스킬로 업무 자동화하는 법

Summary: Claude Chat → Claude Co-working → Claude Code (repeatable automation workflow)

• Goal: Lee Jung-bu explains how to combine Claude Chat, Claude Co-working, and Claude Code into a single repeatable workflow to automate research + content production with higher efficiency.
• Core idea: Use Projects (Chat) to store instructions/memory/files → import them into Co-working to process local files using connectors/skills → export packaged outputs and run them in Claude Code via plugins/commands.

---

1) Claude Chat: Projects (instructions + memory + references)

Setup

• Install the desktop app (Windows “Windows Useful Claude” / Mac app for MacBook).

Project structure

A Project is the center of the workflow. It includes:

• Memory: conversation history inside the project (speaker notes you typically don’t need to edit it).
• Instructions / References: detailed instructions, optionally appended as references.
• Files: uploaded/attached material used later.

Example: YouTube “marketing team” project

• The speaker demonstrates creating a project with:
  • Brand identity
  • Roles
  • Content guidelines
  • Work methods
• After the instructions work, Claude can generate outputs like a 10-month step-by-step roadmap (example goal: “reach 100,000 subscribers within 10 months”).

---

2) Claude Co-working: local-folder automation + multi-agent tasks

What you do

• In Co-working, you select a local folder (e.g., manuscript folder + card-news images/scripts).
• Co-working loads the earlier Project and binds its instructions/files into the workspace.

Multi-agent task pattern

The speaker uses multiple tasks, including parallel work:

• Task 1: Manuscript analysis agent (content marketer)
  • Analyzes the manuscript folder
  • Produces 5 card-news slides
• Task 2: Card news designer agent
  • Uses Task 1 results + templates
  • Generates card news that matches the template

Connectors/skills mentioned

• The workflow uses skills and at least one external integration for image generation.
• Mentions:
  • Nanobara / MCP for generating images
  • Docs for producing a Word document output

Outputs saved locally

Co-working produces:

• A structured Word/Docs file (e.g., a five-page card-news format script)
• Card news results generated using the template + connector-driven image generation

Advantage highlighted: combining local files + connectors/skills and saving deliverables back locally.

---

3) Connectors, Skills, and MCP (concepts explained)

• MCP connector: brings in and uses external components/resources inside Claude
  • API costs may apply (speaker notes this).
• Skills: “manuals/recipes” for repetitive tasks.
• Team/agent model analogy:
  • Project → reusable instruction bundle
  • Connector → external integrations
  • Skill → reusable workflow “recipe”
  • Sub-agents/agent teams coordinate task execution

Skill creation tutorial: “Skill Creator”

• Demonstrates creating a new skill (example: “Card News Maker”) based on the working Co-working workflow.
• Registration/publishing workflow:
  1. Export by compressing to G-file / MD
  2. Upload via Settings → Features → Add
  3. Later invoke it by name in future runs

---

4) More skill creation: web/internet research + drafting scripts

Example: trend research skill

• Creates a skill for trend research using:
  • Web browsing / Web Search connector (Claude 4)
• Produces a YouTube script draft, e.g.:
  • “Top 3 AI Tools to Watch Out For”
• Uses a provided Claude MD file as the instruction template.

Parameter refinement

• Example adjustment: search in batches (mentions “search ten items at a time”).
• Draft length targets about 10 minutes (typical video duration).

Export/registration steps (repeat pattern)

• Create via Skill Creator
• Export compressed file
• Upload/enable via settings

---

5) Claude Code: plugins + Flat Mode planning/testing + team usability

Using Claude Code with the same workflow

• The speaker continues using the same folder and manages permissions.

Modes mentioned

• Plan Mode: recommended for safety when something important is at stake.
• Flat Mode: convenient for testing/modifying structure before running.

Tutorial: build an automation plugin

• The speaker demonstrates packaging skills as a plugin (example name: “Professor Lee YouTube Marketer”).
• Plugin commands include:
  • Reference skills from folders like AI Trend Research and Card News Generator
  • Generate a file structure using manifest JSON
  • Save into a designated work file/workspace

Testing and deployment flow

• Run in Flat Mode test
• Approve plans/permissions when prompted
• Upload the generated plugin G-file
• Use plugin commands for automation with teammates

---

6) End-to-end demo: plugin executes multiple steps

What the plugin runs

• Web search (trend tools) → create manuscripts/drafts
• Card-news generation from the manuscript via the card-news skill

Issue reported + workaround

• When generating card news for different tools, the output format was inconsistent with the intended template.
• Workaround: manually align the script to one “good” observed format (e.g., matching Gamma’s structure).

Result

• Tasks completed successfully with many pages generated (example: 15 pages = 3 tools × 5 cards/tool).
• Expected follow-up: later correction/modification.

---

Key claims / version note

• Mentions “Son(e) 4.6” / context updates:
  • Claude Code context window increased from ~200k tokens to ~1 million tokens
  • Claimed benefit: handle more/longer documents

---

Main speakers / sources

• Main speaker: Lee Jung-bu (an AI professor)
• Referenced channel/content source: “AI Professor Lee Jong-beom” (YouTube)