geoff-ji/jiharn/news-digest

news-digest

Trigger

Use this skill when the user asks for a one-off or scheduled digest of news in a topic (tech, finance, sports, a specific company, ...). When a cron job fires with skill_hint=news-digest, this skill is loaded automatically.

Do NOT use this for:

• Single-article summaries → web-summary.
• Academic papers → daily-paper-digest.
• Verifying a specific claim → direct web_search then read_url.

Inputs

• topic: subject area (e.g. "AI 创业", "OpenAI", "电动车"). If

missing, ask once instead of inventing one.

• window: time window — default "last 24h" for daily, "last 7d" for

weekly.

• max_items: default 5, hard cap 8.
• region: optional geographic / language hint (e.g. "中文媒体").

Steps

1. web_search for <topic> filtered to the window. Prefer reputable

outlets, drop obvious aggregators / SEO farms.

1. Dedup by canonical URL, then by headline-similarity > 0.8.
2. For the top items, read_url only when the headline alone isn't

enough (e.g. ambiguous, possible clickbait). Skip when it is.

1. For each item produce: headline, source domain, one-line takeaway.
2. Order newest first. Keep the whole reply IM-sized — no per-item

markdown headings, no quote dumps.

Verification

• Every item carries a real URL.
• Headlines and dates are not invented.
• If fewer than max_items quality items exist, return the smaller

count and say so — do not pad.

• For scheduled runs, only return [SILENT] when the topic itself is

unresolved.

Failure Signals

• topic missing and not recoverable.
• All search results blocked / paywalled.
• Same story repeats across the digest (dedup is failing).
• Output reads like raw search results rather than a curated digest.