cody1991/culture-trend-curator

Culture Trend Curator

Purpose

Curate public-trend-based recommendations for books, movies, and TV shows. Prioritize recent cultural heat, reputable top lists, long-term ratings, rating-count confidence, and category diversity over personal-profile matching.

Use When

Use this skill for requests such as:

• Recommend books or films based on recent popularity and public lists.
• Produce weekly/monthly/yearly book-and-film digests.
• Compare recent hot items with long-term classics.
• Create scheduled recommendation reports or notification content.
• Avoid relying on the user's personal reading/watch history.

Default Operating Mode

Assume the default digest is weekly unless the user specifies otherwise:

• Time window: last 30 days plus current-year lists.
• Output: 5 books and 5 movies/TV items.
• Books: prioritize Chinese-readable books, especially Chinese editions, translated editions, Chinese publishing/discussion signals, or books easily readable by a Chinese-language audience. Avoid turning the book section into direct English-list translation.
• Books must not become a new-book list: treat new-book pages as discovery sources only. In a normal weekly digest, mix recent hot books with stable backlist, modern classics, long-term high-score works, and media/community picks.
• Movies/TV: prioritize works that are watchable or actively discussable in Chinese-language contexts. Keep one global film-history classic catch-up slot by default; it does not need to be a Chinese-language work, but it must have widely recognized classic value and convenient Chinese viewing/discussion signals.
• Style: concise editorial curation.
• Personal data: default to public sources; only use personal records when explicitly requested.
• Sources: prefer public rankings, reputable media lists, platform charts, and cross-source consensus.
• Links: preserve clickable source links in the final Markdown whenever a URL is available.
• Artifacts: when the user asks for a publishable output, report, article, or reusable result, create a .md artifact instead of only answering inline.

Workflow

1. Define the curation brief.

   • Identify content types: books, films, TV, or all.
   • Identify time window: 7 days, 30 days, current year, past year, or all-time classics.
   • Use the default locale rule unless overridden: books should be Chinese-readable first; movies/TV should be Chinese-watchable or Chinese-discussable first, with one global classic catch-up slot.

2. Collect candidate items.

   • Read references/source_strategy.md.
   • Search multiple independent sources instead of relying on one platform.
   • Capture title, creator, type, source, URL, rating, rating count, list/rank, year, tags, and why it appeared.
   • Deduplicate translated titles and repeated editions.

3. Rank and filter.

   • Read references/scoring_rules.md.
   • Filter obvious low-quality items: very low rating, very small sample, pure marketing lists with no evidence.
   • Balance recent heat with long-term quality.
   • Enforce diversity: avoid recommending only one genre, country, content type, or source family.
   • For books, cap new-book candidates and explicitly include stable/backlist/classic choices so the result is not just a Douban latest-books digest.

4. Render the digest.

   • Read references/output_format.md.
   • If the user asks to publish, post, send to a platform, or adapt for WeChat Official Account/newsletter/social, read references/publishing_formats.md.
   • Use scripts/rank_items.py when candidates are saved as JSON.
   • Use scripts/render_digest.py to generate Markdown from ranked JSON.
   • Include for each item: recommendation reason, why now, source signal, and risk note.
   • Use clickable Markdown links for item titles, source notes, and final references when URLs are available.
   • Save the final publishable draft as a .md file when the user asks for a concrete artifact or platform-ready article.
   • For configured WeCom delivery, use scripts/send_wecom_notification.sh only after the article and cover exist. Keep the Webhook in ignored local configuration; never put it in skill instructions, artifacts, or Git.

Optional Delivery

When the user asks to send a completed digest to a configured WeCom group, read the delivery instructions in references/publishing_formats.md. The default delivery is a title notification, cover image, and Markdown attachment. Send the full article body only when explicitly requested because it may span several group messages.

5. Save recommendation history when appropriate.
   • If running repeatedly, store emitted titles in a local history JSON file in the caller's project or chosen output directory.
   • Record history by default, but do not use it to exclude or down-rank recommendations unless the user explicitly asks to avoid repeats.
   • Use canonical keys and aliases so translated titles and original titles can deduplicate together when deduplication is enabled.
   • Optional duplicate windows: daily = 30 days, weekly = 8 weeks, monthly = 6 months, classic = 90 days.

Candidate JSON Schema

When using scripts, prepare a JSON file as either a list or an object with items:

{
  "items": [
    {
      "title": "Book or film title",
      "creator": "Author, director, or showrunner",
      "type": "book",
      "url": "https://example.com/item",
      "year": 2026,
      "rating": 8.7,
      "rating_scale": 10,
      "rating_count": 12000,
      "source_count": 3,
      "trend_window": "30d",
      "trend_score": 0.8,
      "rank": 4,
      "tags": ["history", "nonfiction"],
      "sources": ["Douban", "WeRead", "media list"],
      "evidence": ["Appeared in a monthly list", "High rating count"],
      "reason": "Why it is worth recommending",
      "risk": "Possible caveat"
    }
  ]
}

Script Usage

Rank candidates:

python3 scripts/rank_items.py candidates.json --out ranked.json --history recommendation_history.json --mode weekly

The default --duplicate-policy none records history later but does not use history to exclude or down-rank items.

Daily mode with a 30-day strong duplicate window, only when explicitly requested:

python3 scripts/rank_items.py candidates.json --out ranked.json --history recommendation_history.json --mode daily --duplicate-policy exclude

Render Markdown digest and append emitted items to history:

python3 scripts/render_digest.py ranked.json --out digest.md --books 5 --films 5 --history-out recommendation_history.json --mode weekly

Quality Bar

Recommend fewer items rather than filling weak slots. If a candidate only has short-term traffic and no quality evidence, exclude it or label it as a hype watch. If sources conflict, explain the uncertainty briefly.