ask-yt

name: ask-yt description: Ask YouTube's built-in Gemini AI about video content using CDP automation. Use when user provides a YouTube URL and wants to ask questions about the video. Supports multi-turn questioning: open the panel once, ask multiple questions. Triggers on: "YouTube 영상에 대해 질문해줘" + URL, "이 유튜브 영상 요약해줘" + URL, "/ask-yt [URL] [question]", User wants to query YouTube's Ask/질문하기 built-in AI feature

⚠️ Language Policy

• Output to user: KOREAN by default
• Phase instructions: ENGLISH (mandatory)

Available MCP Tools

Phase 0: Chrome CDP Auto-Setup (Automated Fallback)

Trigger this phase automatically when:

• open_ask_panel returns a Chrome connection error, OR
• Starting the skill for the first time

Do NOT ask the user to run commands manually. Execute each step using Bash, and only pause to ask the user for decisions.

Step 0-1: Check if CDP is already running

curl -s http://localhost:9222/json/version 2>/dev/null | python3 -c \
  "import sys,json; d=json.load(sys.stdin); print('connected:', d['Browser'])" 2>/dev/null \
  || echo "not_connected"

• If connected: → CDP is already running. Skip to Phase 2.
• If not_connected → proceed to Step 0-2.

Step 0-2: List available Chrome profiles (run silently, present result to user)

python3 -c "
import json, os
base = os.path.expanduser('~/Library/Application Support/Google/Chrome')
profiles = []
for name in sorted(os.listdir(base)):
    pref = os.path.join(base, name, 'Preferences')
    if os.path.isfile(pref):
        try:
            d = json.load(open(pref))
            email = (d.get('account_info') or [{}])[0].get('email', '(없음)')
            profiles.append((name, email))
            print(f'{name}: {email}')
        except: pass
"

Present the profile list to the user and ask which to use via AskUserQuestion (one option per profile). Default to the profile that looks like their main YouTube account.

Step 0-3: Ask permission to kill Chrome and set up CDP

Use AskUserQuestion to confirm:

• "Chrome을 종료하고 CDP 모드로 재시작할까요?"
• Options: "네, 진행해주세요" / "아니요, 수동으로 할게요"

If user declines → provide manual instructions and stop.

Step 0-4: Kill Chrome

pkill -9 -f "Google Chrome" 2>/dev/null; sleep 1; echo "종료 완료"

Step 0-5: Check if Chrome-CDP directory already exists, ask whether to re-copy

PROFILE="<selected_profile>"
CDP_DIR="$HOME/Library/Application Support/Google/Chrome-CDP"
if [ -d "$CDP_DIR/$PROFILE" ]; then
  echo "exists"
else
  echo "not_exists"
fi

• If exists → Use AskUserQuestion:
  • "기존 Chrome-CDP 프로필이 있습니다. 재사용할까요, 아니면 최신 프로필로 다시 복사할까요?"
  • Options: "재사용 (빠름)" / "다시 복사 (최신 쿠키·세션 반영)"
• If not_exists → proceed to copy automatically.

Step 0-6: Copy profile (if needed)

PROFILE="<selected_profile>"
SRC="$HOME/Library/Application Support/Google/Chrome/$PROFILE"
DST="$HOME/Library/Application Support/Google/Chrome-CDP"

rm -rf "$DST/$PROFILE"
mkdir -p "$DST"
cp -r "$SRC" "$DST/$PROFILE"
cp "$HOME/Library/Application Support/Google/Chrome/Local State" "$DST/" 2>/dev/null
rm -f "$DST/$PROFILE/LOCK"
echo "복사 완료"

Step 0-7: Launch Chrome with CDP in background

PROFILE="<selected_profile>"
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
  --remote-debugging-port=9222 \
  --user-data-dir="$HOME/Library/Application Support/Google/Chrome-CDP" \
  --profile-directory="$PROFILE" \
  --no-first-run \
  2>/dev/null &
echo "Chrome 시작됨 (PID: $!)"

Step 0-8: Verify CDP connection (retry loop, up to 10 attempts)

for i in $(seq 1 10); do
  RESULT=$(curl -s http://localhost:9222/json/version 2>/dev/null | python3 -c \
    "import sys,json; d=json.load(sys.stdin); print(d['Browser'])" 2>/dev/null)
  if [ -n "$RESULT" ]; then
    echo "✅ CDP 연결 성공: $RESULT"
    break
  fi
  echo "⏳ 연결 대기 중... ($i/10)"
  sleep 1
done

if [ -z "$RESULT" ]; then
  echo "❌ CDP 연결 실패"
fi

• On success → proceed to Phase 2.
• On failure after 10 attempts → report error to user with the common errors table below.

Common Errors

Phase 1: Input Collection

Collect the following from the user:

1. YouTube URL — must be https://www.youtube.com/watch?v=...
2. Question(s) — what to ask the YouTube AI

If either is missing, use AskUserQuestion.

Phase 2: Open Panel

Call mcp__plugin_ask-yt_yt__open_ask_panel once per video:

tool: mcp__plugin_ask-yt_yt__open_ask_panel
args:
  url: <YouTube URL>
  cdp_port: 9222      (기본값)
  timeout_ms: 30000   (기본값)

On connection error → immediately run Phase 0 (automated setup), then retry open_ask_panel.

After successful open_ask_panel, initialize the JSONL log file using Bash:

python3 -c "
import re, json, sys
url = sys.argv[1]
video_id = re.search(r'[?&]v=([^&]+)', url).group(1)
filename = f'YT_{video_id}_raw.jsonl'
open(filename, 'w').close()  # truncate/create fresh
print(filename)
" "<YouTube URL>"

Save the returned filename (e.g. YT_EQ-Rnx-k-Ec_raw.jsonl) — use it in every Phase 4 append step.

Phase 3: Complete Video Dissection

Goal: Ask enough questions that someone who reads the final document does NOT need to watch the video. Cover every section, every claim, every example, every demo.

Target: 25–30 questions total. Run autonomously without waiting for user input.

Step 3-1: Structure Mapping (ALWAYS ask these 2 first)

Q1 — Full outline:

이 영상의 전체 목차를 타임스탬프와 함께 알려줘.
주요 섹션, 소주제, 각 섹션에서 다루는 핵심 내용을 빠짐없이 정리해줘.

Q2 — Speaker & context:

발표자는 누구이고, 이 발표의 배경(행사명, 개최 맥락, 청중)은 무엇인지 알려줘.
발표자의 전문성과 이 발표를 하게 된 계기도 포함해줘.

Step 3-2: Section-by-Section Excavation

From the outline in Q1, identify every major section. For EACH section, ask ALL of the following sub-questions (adapt wording to use exact section names):

Sub-question A — Verbatim detail:

[섹션명] 부분에서 발표자가 정확히 어떤 말을 했는지 최대한 자세히 알려줘.
구체적인 발언, 예시, 비유, 스토리를 모두 포함해서 설명해줘.

Sub-question B — Data & evidence:

[섹션명]에서 언급된 구체적인 수치, 통계, 사례, 실험 결과가 있으면 전부 알려줘.

Sub-question C — Demos & visuals (if applicable):

[섹션명]에서 화면에 보여준 시연, 코드, 다이어그램, 또는 실시간 데모가 있었으면 설명해줘.

Skip Sub-question C if the section clearly has no demo/visual component.

Step 3-3: Cross-Cutting Deep Dives

After all sections are covered, ask these regardless of topic:

Terminology sweep:

이 영상에서 발표자가 새롭게 정의하거나 독특하게 사용한 용어나 개념이 있어?
각각 발표자의 정의를 그대로 설명해줘.

Most controversial claim:

이 영상에서 가장 논쟁적이거나 반직관적인 주장은 무엇이고,
발표자는 그 근거로 무엇을 제시했어?

Audience Q&A (if exists):

발표 후 청중 질문이나 Q&A 세션이 있었어? 있었다면 어떤 질문과 답변이 오갔는지 알려줘.

Completeness check:

지금까지 내가 묻지 않은 중요한 내용이 이 영상에 남아 있어?
있다면 무엇인지 알려줘.

Step 3-4: Chain Follow-ups (run after EVERY answer throughout)

After each answer, scan for these and immediately ask if found:

Rules:

• Use EXACT quotes and terms from the previous answer — never ask generically
• Never re-ask something already answered
• Chain follow-ups take priority over moving to the next section if an answer raises something significant

Step 3-5: Termination

Stop when:

• All sections from the outline are excavated AND cross-cutting questions are done, OR
• Hard limit: 30 questions

On termination → call close_session → proceed to Phase 5.

Phase 4: Display, Append & Loop

After each ask_video call:

Step 4-1: Show answer to user

**[Q{n}]** {question}

{answer}

Step 4-2: Append to JSONL immediately

Run this Bash command right after displaying the answer (replace {n}, {question}, {answer}, {chips_json}, {jsonl_file} with actual values):

python3 - <<'PYEOF'
import json

entry = {
    "q_num": {n},
    "question": {question_as_json_string},
    "answer": {answer_as_json_string},
    "chips": {chips_as_json_array}
}
with open("{jsonl_file}", "a", encoding="utf-8") as f:
    f.write(json.dumps(entry, ensure_ascii=False) + "\n")
print(f"Q{n} 저장 완료")
PYEOF

• {question_as_json_string}: the question string wrapped in json.dumps() style (properly escaped)
• {answer_as_json_string}: the answer string properly escaped
• {jsonl_file}: the filename initialized in Phase 2

Step 4-3: Loop

Immediately return to Phase 3 (Step 3-2 or 3-4) to determine the next question. Do NOT wait for user input. Keep the chain running autonomously until the termination condition.

Phase 5: Generate Complete Video Report

After closing the session, read the JSONL log and write a document comprehensive enough to fully replace watching the video.

Step 5-1: Read all Q&A from JSONL

python3 -c "
import json
with open('{jsonl_file}', encoding='utf-8') as f:
    entries = [json.loads(line) for line in f if line.strip()]
print(f'총 {len(entries)}개 Q&A 로드됨')
for e in entries:
    print(f\"Q{e['q_num']}: {e['question'][:60]}...\")
"

Verify all entries are present. If any are missing, re-ask those questions via ask_video and append manually.

Step 5-2: Determine report filename

python3 -c "
import re, sys
url = sys.argv[1]
video_id = re.search(r'[?&]v=([^&]+)', url).group(1)
print(f'YT_{video_id}_REPORT.md')
" "<YouTube URL>"

Step 5-3: Write the document

Use the Write tool to create {report_filename} in the current working directory. Use ALL entries from the JSONL — never truncate or summarize Q&A content.

문서 형식:

# {영상 제목} — 완전 해설

> **영상**: {url}
> **발표자**: {speaker name & title}
> **행사**: {event name & date if known}
> **작성일**: {YYYY-MM-DD}
> **분석 질문 수**: {n}개

---

## 한눈에 보기 (Executive Summary)

{영상 전체를 3-5 단락으로 압축. 발표의 핵심 주장, 근거, 결론을 포함.
이 섹션만 읽어도 발표의 80%를 파악할 수 있어야 함.}

---

## 목차

{Q1에서 얻은 타임스탬프 기반 목차를 그대로 재현}

---

## 발표자 & 배경

{Q2 내용. 발표자 소개, 행사 맥락, 발표 동기}

---

## 섹션별 상세 해설

### 1. {섹션명} ({timestamp 범위})

{해당 섹션의 Sub-question A~C 답변을 통합하여 서술.
발표자의 발언을 최대한 재현하되 읽기 쉽게 구조화.
구체적 수치, 예시, 비유를 모두 포함.
데모가 있었으면 "[시연]" 블록으로 표시.}

> **[시연]** {데모 내용 설명}

### 2. {섹션명} ({timestamp 범위})

{...}

(모든 섹션 빠짐없이 포함)

---

## 핵심 용어 사전

| 용어 | 발표자의 정의 |
|------|-------------|
| {term 1} | {definition} |
| {term 2} | {definition} |

---

## 가장 논쟁적인 주장

> **주장**: {controversial claim 원문}

**발표자의 근거:**
{evidence provided}

**비판적 고려:**
{limitations or counterpoints surfaced during questioning}

---

## Q&A 세션 요약

{청중 질문이 있었으면 Q&A 형식으로 재현. 없으면 이 섹션 생략.}

---

## 실전 적용 가이드

### 즉시 실행 (오늘부터)
- {액션 1}
- {액션 2}
- {액션 3}

### 중기 변화 (1-6개월)
- {변화 1}
- {변화 2}

### 큰 방향성 (장기)
- {방향 1}

---

## 핵심 명언

> "{exact quote 1}" — {speaker}

> "{exact quote 2}" — {speaker}

---

## 남은 질문들

{영상에서 다루지 않았거나 답하지 않은 중요한 질문들. 시청자가 추가로 탐구할 주제.}

Rules:

• 섹션별 상세 해설: 발표자의 실제 발언을 최대한 재현 — 요약이 아니라 재현
• 수치·사례·비유 전부 포함, 절대 생략하지 말 것
• Chip 제안은 문서에 포함하지 않음 (UI artifact)
• 질문이 3개 미만이면 이 포맷 대신 단순 Q&A 나열로 대체

Step 5-3: Notify user

Tell the user: 📄 영상 완전 해설 문서가 저장되었습니다: {filename}