YouTube Research → Chinese Brief

Purpose

Convert a research direction into a structured Chinese-language brief backed
by the top 5 authoritative English YouTube talks on that topic.

The output is two-layered:

1. Human-readable — analysis.md with comparison tables, taxonomy, and
   a per-video applicability assessment against a user-supplied application
   context (e.g. 网络测绘, 客服 Agent, 代码评审).
2. Machine-readable — videos.json, findings.json, applicability.json
   that load into SQLite in one line.

The skill bakes in every environment fix required to make yt-dlp actually
work on this macOS + WorkBuddy host (see references/pitfalls.md).

When to trigger

Trigger phrases (Chinese or English):

• "研究 X 方向，搜 YouTube 视频翻译给我"
• "找 X 主题最好的英文讲座"
• "做一份 X 调研"
• "用 YouTube 视频帮我做技术调研"
• "Research X topic from top YouTube talks and translate to Chinese"
• "Build a brief on X from YouTube videos"

Hard preconditions (must be satisfied by user before invocation):

• A research direction (≥3 words).
• Optional but recommended: an application context, so the applicability
  section is grounded. Without it, omit the applicability table entirely.

Workflow

The skill is a strict pipeline. Do not parallelize across stages; the
output of each stage feeds the next.

Stage 1 — Discover (use WebSearch)

Find the top 5 most authoritative English-language YouTube talks on the
research topic. Use 2-3 distinct search queries (broad + specific) and
collect up to 8 candidate IDs, then rank by:

1. Presenter / channel authority (Google, Anthropic, MongoDB, Mem0, Letta
   > individual creators for technical topics).
2. Recency (prefer 2024+ when relevant).
3. Depth (longer > shorter when depth is the user's goal).

Capture: youtube_id, title, channel, url.

Stage 2 — Download subtitles (use scripts/yt_dlp_download.sh)

mkdir -p <output_dir>/vtt
cd <output_dir>/vtt
~/.workbuddy/skills/youtube-research-cn/scripts/yt_dlp_download.sh \
  "https://www.youtube.com/watch?v=<id1>" \
  "https://www.youtube.com/watch?v=<id2>" \
  ...

The script handles every pitfall documented in references/pitfalls.md:
env-var scrubbing, managed Node path, EJS challenge solver, player_client
fallback. If you re-implement the command inline, you MUST carry every
flag from references/pitfalls.md §TL;DR.

After download, verify with ls *.vtt and word-count each file with
wc -w. If a video's VTT is empty or <500 words, the auto-sub failed;
retry once with the relaxed fallback flags the script applies, or drop
the video from the shortlist.

Stage 3 — Parse (use scripts/parse_vtt.py)

~/.workbuddy/skills/youtube-research-cn/scripts/parse_vtt.py <output_dir>/vtt

This produces *.txt siblings next to every *.vtt. Expected output for
a 25-minute talk: 2-7k words after dedup, depending on speech rate.

Stage 4 — Analyze & translate

Read each *.txt in full. For each video, extract:

• Metadata: presenter, source, duration class.
• Core thesis: 1-sentence Chinese summary of the speaker's position.
• Concept taxonomy: which memory types / patterns / evaluations they
  mention (cross-reference against the master taxonomy list in
  references/output_schema.md).
• Architecture dimension coverage: which of {storage, retrieval,
  write path, read path, multi-agent, async, forgetting model} the talk
  actually addresses.

Stage 5 — Compose analysis.md

Follow the structure in references/output_schema.md §analysis.md structure exactly. All tables must be Markdown. The applicability
section is only generated if the user supplied an application context.

Stage 6 — Produce SQLite-ready JSON + schema.sql

Emit four files in <output_dir>:

• videos.json — flat array, one object per video (Table 1 schema).
• findings.json — flat array, one object per (video, dimension, key)
  triple (Table 2 schema). For this skill the file is named architecture.json
  to reflect that architecture facts (core_abstraction, read_path, write_path,
  storage, forgetting_model, multi_agent, async) are the primary findings.
• applicability.json — only if application context was supplied
  (Table 3 schema).
• memory_types.json — master dictionary of memory categories (Table 2'
  schema).
• schema.sql — CREATE TABLE for all four + a v_top_recommendations
  convenience view.

The user can load all four files into SQLite with one command after applying
the schema:

sqlite3 research.db < schema.sql
~/.workbuddy/skills/youtube-research-cn/scripts/load_to_sqlite.py research.db .

Resources

This skill ships three reusable artifacts:

scripts/

references/

Notes

• The skill deliberately separates discovery (WebSearch — judgement
  needed) from download (deterministic script) from analysis
  (judgement). Don't try to fully automate analysis — model judgement is
  the whole point.
• The skill is language-agnostic for input but the output is always
  Chinese unless the user explicitly says otherwise.
• This skill does not generate original commentary about the research
  topic. It summarizes and structures what the videos actually said.
  Cross-video synthesis is allowed but must be flagged as such.
• If a video's content overlaps ≥80% with another already-downloaded
  video, prefer the shorter one to save analysis time, but note the
  overlap in the analysis.