코드 한 줄도 안 짜고 마케팅 에이전트 만든 방법(딸깍을 잘하는 방법)

Core idea (how the “copy + modify” approach works)

The video explains a practical, automation-driven way to build a marketing agent (without manually writing much code) by leveraging Claude Code, LLM prompting, and social media publishing automation.

The creator frames AI’s strength as:

Copy what already works well, then modify it to fit your needs.

They base the workflow on an existing tutorial by Builder Josh, then reproduce and adapt it.

Referenced prior tutorial / method being copied

Builder Josh’s marketing agent workflow includes:

• Trend analysis
• Card news / image generation
• Automatic social media publishing

Implementation mentions:

• Jemina’s API (as stated in subtitles; the name may be mis-transcribed)
• Automation tools (general)

Step-by-step guide (main tutorial portion)

1. Get the source material

• Copy the Builder Josh video URL + script.
• Use a Chrome extension like “YouTube Summary” to extract the script.
• Alternative: provide the LM/video link so subtitles/summary appear together.

2. Use Claude Code to generate project scaffolding

• Paste the Builder Josh script + URL into Claude Code.
• Prompt Claude to generate:
  • A task list in an MD file (with the user’s brand)
  • Also generate HTML for human readability

3. Why generate both MD and HTML

• Markdown (MD): structured, readable lines for the AI workflow.
• HTML: more visually digestible (colors/images) for humans at a glance.

The creator cites an industry rationale:

• An Entropic Claude Code team engineer (Tarik Siyppa, per subtitles) argues they don’t read Markdown over ~100 lines and shifted toward HTML.
• Another authority (Kapathy / Kapasi, as transcribed) agrees that requesting HTML formatting helps structure outputs.

4. Prepare exactly three manual inputs before running the workflow

The user must provide:

1. Brand information

   • Company name, website URL, main services, target customers, etc.

2. API key

   • Described as an access “password.”
   • Tip: use Claude’s PlayWrite MCP to guide the user to the API key issuance page and re-register it into the MV file/task context.

3. Reference/design assets

   • Pinterest for card-news/frame design references
   • GetDesign (via GetDesign.md files) with design systems from brands (e.g., Apple)
   • “Umind C” (as transcribed) containing 6,000+ GPT image prompts
   • Mentions a model code name: “Duct Tape” for the image prompt/generation reference

5. Run tasks sequentially via Claude Code

• Claude processes the task list in order.
• The creator warns you may get stuck at least once and recommends debugging by:
  • Step 1: paste the error message exactly
  • Step 2: request step-by-step debugging (iterative diagnosis as implied)

6. Publishing automation (eliminate manual posting)

To avoid repeatedly uploading posts manually:

• Use Buffer as the social media scheduling tool.
• Claude is asked to handle:
  • Buffer API key onboarding
  • Agent integration
  • Scheduling reservations based on time/day/platform automatically

7. Final deliverable: HTML project overview

After completion, Claude outputs an HTML document to understand the project as a whole.

Claimed outcome / efficiency gain

• The process reduces time to build a marketing agent from about a week to a day.
• Core philosophy:

The challenge isn’t “making AI work”—it’s creating the right structure so people/agents can execute tasks together.

Main speakers / sources mentioned

• Builder Josh (video/tutorial being copied)
• Tarik Siyppa (Entropic Claude Code team engineer; cited about switching toward HTML)
• Kapathy / Kapasi (described as a top AI authority/co-founder of OpenAI and Tesla head of AI—name transcribed from subtitles)
• The video creator / presenter (author of the workflow)
• Tools/companies referenced:
  • Entropic Claude Code, Claude, Buffer, Pinterest, GetDesign (design system repo), MCP (PlayWrite)