💬 Don't read 1,800 comments. Answer the 8–10 that matter — today.

Paste a video link and the skill collects the top 500 comments by likes, analyzes them, and hands you a report with the 8–10 comments you should reply to first, plus a ready-to-edit draft reply for each.

• One conversation produces a 6-section report: sentiment breakdown → top repeated questions → content ideas → complaint list → reply priorities → best comments.
• Every finding quotes the original comment (username + like count) verbatim, so you can verify the report line-by-line against the raw CSV.
• No YouTube API key, no account login. Collection runs on free, open-source yt-dlp, and all data stays on your machine.

📦 Real content preview

This is a new listing with no reviews yet, so judge it by the actual content. The passages below are excerpted directly from the analysis-criteria document included in this skill (references/analysis-framework.md), translated from the Korean source.

Excerpt 1 — Reply-priority rules. Not just "most likes first" — priorities weigh reputation risk and reply visibility.

Scoring method (qualitative rules):

• Tier 1: high-likes + unanswered questions, and strong negativity + many likes/replies (reputation risk)
• Tier 2: representative questions from FAQ clusters; improvement requests that are concrete and actionable
• Tier 3: top-liked thank-you/praise comments (short replies to maintain the relationship)
• Excluded: spam/ads, bare tags, bad-faith provocation (flagged "no response recommended" instead of a draft)

Excerpt 2 — Intent normalization in FAQ clustering. Questions phrased differently but solved by the same answer are merged into one cluster — the part keyword-matching tools can't do.

2. Intent normalization: normalize by intent, not wording.
   e.g. "Can I use an air fryer?" / "No oven — any alternative?" / "Would a microwave work?"
   → intent: "whether alternative cooking equipment works".
3. Cluster-merge criterion: if one answer resolves them all, same cluster.
   If the answers must differ, split them (prevents over-merging).

Excerpt 3 — Draft-writing rules. The goal is a reply that's one edit away from posting, not a copy-paste template.

• 2–4 sentences; reference the comment's specifics so it doesn't read like a paste job.
• Anything needing fact-checking (price, schedule, specs) is marked with a [verify: ...] placeholder.
• Drafts for negative comments acknowledge what's valid and mention an improvement plan — never a defensive rebuttal.

Beyond these, the same document covers a sentiment-labeling rubric (questions default to neutral, mixed-sentiment handling, a hold-judgment rule for sarcasm and memes), best-comment selection criteria, and channel-level cross-analysis — 5 sections in total.

🗂️ What's included

What each run produces, per video:

• report-{videoID}.md — summary → sentiment breakdown → top repeated questions (5–10) → content ideas → complaints & improvement requests → top 8–10 reply priorities (with drafts) → ~5 best comments
• comments-{videoID}.csv — 10 columns: author, like count, text, reply count, is-reply flag, parent ID, comment ID, timestamp, uploader-replied flag, hearted flag. Opens cleanly in Excel (UTF-8 BOM, no mojibake)
• channel-report-{date}.md — with 2+ links, adds a cross-video report: per-video sentiment comparison table plus channel-wide request/complaint patterns ("overheating mentioned — 3 of 4 videos")

Sampling coverage is always stated at the top of the report, e.g. "Analyzed top 500 of 1,842 comments by likes — 27% coverage."

💬 Three ways people use it

1. Creator — the morning after an upload

"Analyze the comments on yesterday's video: https://youtube.com/watch?v=XXXX — make the reply drafts casual, with emoji."

→ You get the sentiment breakdown and 3 priority actions in chat first, then the top 8–10 reply priorities in a report file, each with a draft matched to your channel's voice. Spam and provocation get "no response recommended" instead of a draft.

2. Brand marketer — sentiment report on a 4-video campaign

"Here are 4 links from our campaign. Compare reactions per video and flag any complaints that show up across them."

→ A cross-video report with a per-video sentiment comparison table, plus requests and complaints that recur in 2+ videos aggregated as "channel-wide patterns." Every item quotes its source comment, ready to drop into a client report.

3. Online instructor — building a pinned FAQ comment

"Pull the repeated questions from this course video's comments and draft a pinned FAQ comment."

→ Questions with the same intent — however they're phrased — merge into one cluster, presented with frequency, total likes, and representative quotes. The list doubles as pinned-comment material and curriculum feedback.

✅ Buy this if you are

• A YouTube creator with hundreds of comments per video and the recurring question "where do I even start replying?"
• A marketing/PR/channel manager who reports audience sentiment across videos, with source quotes required
• An online instructor mining repeated questions to improve FAQs and curriculum
• Already using Claude Code and want comment review done in one conversation

⛔ Skip it if you

• Want auto-posted replies — this skill stops at drafts; you post them yourself (no account auth, no automation, so nothing conflicts with channel policies)
• Need exhaustive analysis of every comment — it samples the top N by likes (default 500, adjustable), and coverage is always disclosed in the report
• Expect real-time monitoring or alerts — each run is a snapshot at execution time
• Run a channel primarily in languages other than Korean or English — sentiment analysis is tuned for those two, and sarcasm/memes can be misread, so held-judgment counts are reported separately
• Don't use Claude Code (or a Claude agent environment)

🔍 FAQ

Q1. How do I receive it and use it?
Immediately after purchase you download the skill files (SKILL.md + analysis-criteria document + collection script). Drop the folder into your Claude Code skills directory — that's the whole install. Then paste a YouTube link in chat and ask for a comment analysis; the skill activates automatically. The only prerequisite is the free yt-dlp tool (pip install yt-dlp or brew install yt-dlp).

Q2. What about language? The skill files are in Korean.
The skill's internal instructions and analysis criteria are written in Korean — Claude reads them natively, and the skill directs the AI to respond in whatever language you chat in, translating criteria and examples on the fly. As a tool, it fetches comments in any language yt-dlp can access; sentiment analysis is tuned for Korean and English, so treat results on other languages as approximate. Your reports and reply drafts come out in your conversation language.

Q3. How do updates work?
When the analysis criteria or the script improve, the new version is published to this listing. Buyers re-download at no extra cost.

Q4. Can I get a refund?
This is a digital download; refunds follow the policy of the platform you purchased on (Capafy).

Q5. Do I need a separate AI subscription?
This is a skill that runs inside Claude Code, so you need a working Claude Code environment. Beyond that: no YouTube API key, no account login, no paid services. yt-dlp and Python 3.9+ are free.

Q6. Can I use it commercially?
Yes. Use the reports and reply drafts for channel operations, client reporting, course management, and other commercial purposes. Reselling or redistributing the skill files themselves is not permitted.

🛒 Purchase info

$3 · one-time payment · unlimited use. Analyze as many videos as you like at no extra cost. If comment review has been costing you 1–2 hours per video, the first analysis pays for itself.