| name | news |
| description | Curate a Markdown news digest for a specified domain — tech (software engineering / LLMs), global (world politics, economy, society, academia/science), or Japan (Japan-domestic stories summarized in English from Japanese-language sources). Use when the user requests a news digest, briefing, or summary for any of these domains, or when invoked by a scheduler to produce a periodic news update. |
| license | MIT |
news
Produces a curated Markdown digest of recent news in a specified domain. The skill collects, filters, and synthesizes — it does not handle scheduling, transport, or delivery. The caller (a scheduler, a wrapper, the user directly) is responsible for invocation cadence, prior-digest history (for dedup), and any downstream posting.
The shared rules below apply to every domain. Domain-specific sources, coverage, skip rules, and output-language guidance live in per-domain reference files.
Domains
| Domain | Reference file | What it covers |
|---|
tech | tech.md | Software engineering in the LLM era — model releases, research, developer tooling, practical engineering. |
global | global.md | World politics, economy, society, academia/science with explicit cross-perspective sourcing. |
japan | japan.md | Japan-domestic stories, summarized in English from Japanese-language sources. |
When invoked, the caller specifies which domain ("tech", "global", or "japan"). Read the matching reference file before producing the digest — it lists the sources to use, the editorial focus, what to skip, and any domain-specific quality rules.
Volume
- Aim for 5–10 items per digest.
- If there is genuinely nothing worthwhile in the period, return fewer items — even 1–2, or a short "quiet period, nothing notable" note. Never pad to hit a count.
- A small number of strong items beats a long list of mediocre ones.
Item format — mixed depth
Each digest mixes two depths:
- Major items (the lead 2–4 stories): bold headline as a link to the source, then a 3–5 sentence paragraph covering:
- What happened.
- Why it matters — the implication, stake, or change in trajectory.
- Notable counterpoints / how different sides frame it if the story is contested.
- Outlet name on a short trailing line, e.g.
_Source: Reuters_. Include the X handle if the source is X.
- Minor items (the rest): one-line headline + outlet + link, as bullets under a
## Other items heading.
Lead with major items.
Output shape
**<Domain> digest — <YYYY-MM-DD>**
### [Headline as link](https://example.com/article)
3–5 sentence summary: what happened, why it matters, any counterpoint or framing note worth including.
_Source: Outlet name_
### [Next major headline](...)
...
## Other items
- [Headline](url) — Outlet
- [Headline](url) — Outlet
The exact header label comes from the per-domain reference file (Tech digest, Global digest, Japan digest). The caller is responsible for any downstream chunking, embedding, or formatting required by the delivery channel.
Citation discipline
Every factual claim has an inline source. If a claim cannot be supported by a source you have actually verified, omit it rather than filling the gap. "Not enough information to summarize confidently" is a valid output. Never invent quotes, statistics, attributions, outlet names, or links — every URL must point to a real, fetchable page.
Dedup — caller-supplied
If the caller provides a list of recently-covered headlines or URLs (typically from prior digests in some recent window — the caller decides the window), exclude those stories from this digest. Re-cover a previously-included story only when there is a meaningful new development, and lead the summary with what changed.
If no prior list is provided, do not maintain or invent one — just produce the best digest from the current period's sources.
Quality checks before emitting
Run these before producing the final output. If a check fails, fix it (drop an item, swap a source, add a counterpoint) before emitting.
- Diverse sources — not all items from one outlet or one viewpoint. If 4 of 6 items come from the same publisher, that is a smell.
- Standard-outlet anchor — every major item has at least one standard/established outlet or official channel as a source, not just X.
- X verification — anything sourced from X has been sanity-checked against another source. If unverifiable, drop it (or label clearly as "unverified, via @handle" — and only do this for genuinely interesting tips).
- Contested-story framing — if a story is genuinely contested, the summary briefly notes how different sides frame it. One sentence is enough.
- No fabricated sources — every link goes to a real page; outlet names are real.
The per-domain reference file may add domain-specific quality rules (e.g., cross-cultural sampling for global, Bunshun-verification for japan). Apply those in addition to the shared checks above.
Tone
Calm, factual, low on hype. Match the voice of a careful editor briefing a peer. Avoid editorialization in the summary itself — let the contested-story framing note do that work where applicable.
Output language
Domain-dependent — see the per-domain reference file. (For tech and global the output is English; for japan the output is also English even though the sources are primarily Japanese — see japan.md for translation discipline.)
