1개월 만에 조회수 300만...AI로 20분 영상 만들기? 이건 진짜 혁명입니다(유튜버들 비상 걸렸네요)

Key technological concepts & claims

• YouTube/AI-driven moderation & duplication filtering: The speaker argues that many AI videos are being deleted because platforms can detect duplicate/similar content more effectively as AI learns patterns in YouTube’s recommendation system. After enough data accumulates, AI can pre-filter items as “potential duplicates.”

• Shift from copying to “benchmarking” (structured competitive analysis): Instead of copying other videos or scripts directly, the speaker recommends benchmarking overseas content first, then adapting it for local audiences.

• Script benchmarking workflow using LLM tools (NotebookLM + Google Docs + Drive):

  • Don’t just paste a list of YouTube links into an AI tool. The speaker says that approach is “useless/false” because it doesn’t analyze the full video/script content holistically.
  • Create a Google Docs blank document, upload/select materials via Google Drive, and use NotebookLM to analyze the drive content.
  • The AI should analyze structure, characteristics, strengths, and generate a benchmarking guide.

• Benchmark structure template (“5 parts/visions”): The speaker repeatedly references a common script template consisting of:

  1. Structure analysis
  2. Hook diagnosis
  3. Cause analysis
  4. Solution
  5. Vision/ending

• Major warning: template homogeneity leads to removals: If creators copy the same structure from famous channels without meaningful changes, scripts across channels become too similar—contributing to deletions.

• Persona (“persona assignment”) as the main differentiator: The speaker emphasizes that scripts must change not only by topic/materials, but also by persona (e.g., a psychology angle for people in their 20s vs. 30s). They claim persona adaptation adjusts the script appropriately even when raw content overlaps.

• Solution section risk (hallucination / weak reasoning):

  • The speaker claims AI often refuses or avoids “risk-taking” reasoning, producing generic/“floating” solutions.
  • They advise paying special attention to the solution portion since it’s the hardest to make accurate and non-generic.

• Avoid “fake expansion” / hallucination: The speaker criticizes trying to extend a short AI draft by prompting it to “make it longer” without real research, because the model may invent details (“hallucination”). This harms quality and increases controversy risk (which is also tied to deletion).

• Deep research requirement: For reliable writing, the speaker argues you must do deep research in the tool (not just a short prompt). They also mention providing guidelines along with research prompts to improve output that is easier to reorganize.

---

Product features / tools mentioned

• NotebookLM (used after copying/uploading scripts/articles into Google Drive / Google Docs) to:

  • analyze script structure
  • generate a benchmarking guide
  • output structured sections that can be reorganized and rewritten

• Google Docs / Google Drive

  • used as the workspace and input source so the AI can analyze materials
  • upload/select files from Drive

• Flux 1 model / Nanobana Mentioned as options for generating or making assets/scripts/images; exact usage is ambiguous in the subtitles.

• AutoWhisk browser extension / “Whisk” tool

  • an automation tool to extract image proportions and generate images automatically
  • used for semi-automation: assembling subtitles + images + effects

• Custom “automation program/site” (speaker’s PG “Money Monster channel” workflow)

  • generates a full video pipeline from a script, including:
    • subtitles
    • audio preview/voice
    • fade-in/out, zoom effects
    • image generation and timeline assembly
  • claim: real automation is only when the full pipeline runs end-to-end (not just image generation)

---

Tutorial-like workflow (end-to-end)

1. Benchmark overseas first (e.g., psychology/economics/history categories) before copying local trends.
2. Collect competitor scripts/materials (around 20 items), but avoid naive “link-pasting.”
3. Use NotebookLM with uploaded scripts to generate:
   • structure analysis
   • hook diagnosis
   • cause analysis
   • solutions
   • vision/ending
4. Rewrite/adapt using persona changes:
   • modify persona-dependent sections
   • tweak the solution more carefully to avoid generic or hallucinated answers
5. Generate images/characters:
   • warn against generic “stickman” characters made mindlessly via mass automation
   • recommend creating at least one distinctive character aligned with the channel theme
6. Generate video:
   • ensure subtitles + images align with the script
   • add editing effects (fade, zoom, etc.)
7. Use automation programs:
   • contrast semi-automation vs full automation and emphasize time savings

---

Review / analysis points

• Meta warning: The speaker frames current YouTube risk as stemming from AI-driven content duplication plus increased issues like controversy/hallucinations, which raise the likelihood of deletion.

• Time/cost argument for automation viability:

  • Claims manual video generation can take ~9 hours.
  • Claims automation reduces this to ~30 minutes (even with adaptation included; sometimes ~5 minutes is mentioned).
  • Includes a rough revenue threshold: to justify 9 hours, they estimate needing ~3.2 million KRW in YouTube revenue (as stated).

• Quality vs automation tradeoff: Automation must preserve brand identity through character design and correct persona/solution editing; otherwise channels “fail” or receive poor exposure.

Core idea: automate the pipeline, but keep the script differentiated (structure + persona + non-generic solutions) to avoid removal risk.

---

Main speakers/sources

• Main speaker: The YouTube creator delivering the lecture (no name given in subtitles). They repeatedly refer to themselves as the instructor/PG for a “Money Monster” channel and use a metaphorical “as a Google… eagle” reference.

• External sources referenced:

  • YouTube/Google (described as introducing AI features that analyze YouTube algorithms)
  • NotebookLM / AutoWhisk / Flux 1 / Nanobana as tools/models