---

name: daily-curator description: >- Generate a curated daily reading list of high-quality articles. Use when the user says "今日推荐", "今天读什么", "daily reads", "morning briefing", "推荐文章", "每日推荐", or invokes "$daily-curator". Also triggers on "本周推荐" (weekly mode) or "只看 AI" (category filter). Does NOT handle RSS subscription management, feed reader setup, or newsletter configuration.

Daily Curator

A curator with taste — picks today's most worth-reading articles from high-quality sources.

Inputs

• mode: daily (default) | weekly
• category: all (default) | ai | engineering | thinking
• count: 5–10 (default 7)
• output_language: user's language, fallback Chinese

Feed Management

Users manage their feeds with simple conversational commands:

• "关注 https://example.com/feed.xml" → append to ~/.daily-curator/feeds.txt
• "取消关注 example.com" → remove matching line from feeds.txt
• "我的信源" / "list feeds" → show current feeds.txt content
• "导入 OPML https://..." → run scripts/import-opml.sh, append all extracted URLs to feeds.txt

The file ~/.daily-curator/feeds.txt is one URL per line, dead simple:

https://simonwillison.net/atom/entries/
https://jvns.ca/atom.xml
https://danluu.com/atom.xml

Lines starting with # are comments. Empty lines are ignored.

Source Resolution Order

1. ~/.daily-curator/feeds.txt — user's personal feeds (primary)
2. references/curated-feeds.md — built-in defaults (fallback if feeds.txt doesn't exist)

When feeds.txt exists, it IS the source list. The curated-feeds.md serves only as a starter template and reference for quality tier classification.

Workflow

Step 1: Load Sources

• If ~/.daily-curator/feeds.txt exists, use it as the source list.
• Otherwise, fall back to extracting URLs from references/curated-feeds.md.
• If the user specified a category filter, cross-reference against references/curated-feeds.md for category tags; user-added feeds without a known category are included in all categories.

Step 2: Fetch Content

• Run scripts/fetch-feeds.sh to batch-fetch all feeds in one pass.
• If the script is unavailable, fall back to WebFetch for each feed URL.
• Collect from each entry: title, URL, published date, summary or description.

Step 3: Score & Filter

• Read references/scoring-and-filtering.md.
• Apply the scoring formula: timeliness × source quality × content depth.
• Deduplicate by canonical URL and fuzzy title similarity.
• If a previous digest exists in conversation context, skip already-recommended items.

Step 4: Select Top N

• Rank by composite score, select the top count items.
• If fewer than 3 items clear the quality bar, say so honestly rather than padding.
• Group by natural topic clusters when applicable.

Step 5: Format Output

• Read references/output-format.md.
• Generate the structured reading list.

Working Rules

• Never pad the list with low-quality items to meet count.
• Prefer depth over breadth — one great article beats three mediocre ones.
• Keep summaries grounded in actual content, not hallucinated.
• If a feed is unreachable, skip it silently (don't error the whole run).
• Same-topic merge: if 2+ articles cover the same event, keep the best and note 另见: SourceB.

Response Shape

• Short preamble: date + theme keywords.
• Numbered list of recommendations.
• Each item: title, source, one-line curator's note, link.
• Footer: source count and quality note.

References

• Read references/curated-feeds.md for the source list.
• Read references/scoring-and-filtering.md for scoring rules.
• Read references/output-format.md for format details and examples.