News

News Skill — AI & Tech Digest

Replaces the "open 12 tabs every morning" habit with one command: a fast, scannable, link-rich brief on what actually shipped.

Customize this skill — this is where the expertise lives. A generic source list gives generic results. Edit the Audience line below and the source list in sources.md to match your field: the outlets you trust, the people you follow, and the tools/topics you want flagged. The specifics are what turn a generalist into an expert.

Audience (edit me): someone who tracks AI and wants the signal without the noise — model releases, pricing, and practical how-to.

Default tone: terse, scannable, link-rich. No preamble, no fluff. The reader skims.

---

Invocation

Input	Behavior
/news	Full digest, last 7 days
/news today	Last 24 hours only
/news week	Last 7 days (default)
/news month	Last 30 days
/news <topic>	Free-form topic (e.g. /news agents, /news Anthropic, /news open models)

Combine freely: /news today agents.

---

Execution flow

1. Parse args. Default window = 7 days.
2. Fan out parallel searches. Issue all WebSearch / WebFetch calls in a single message — never serial. Aim for 12–15 parallel calls per run.
3. Dedupe by headline / URL similarity.
4. Tag each item with: source, date, category, and a ★ flag if it's directly relevant to your stack (the list you define in sources.md).
5. Synthesize into the output format below.

Source buckets to hit in parallel live in sources.md — read it when planning the fan-out. At minimum:

• Official release pages: the labs you care about (Anthropic, OpenAI, Google, etc.).
• Practical / how-to trackers: independent writers who surface what's actually useful.
• Search: "<lab>" release announcement <date> and "<topic>" technique <date>.

---

★ Relevance tagging

Flag an item with ★ if it touches something on your watch-list (defined in sources.md) — a tool you use, a platform you build on, a competitor, a topic you're tracking. For each ★ item, add a one-line What to do: (e.g. "Test the new model on our eval set", "Update our MCP config for the new transport").

This filter is the expertise. It's the difference between "here's all the news" and "here's what matters to you."

---

Output format

Keep it terse. No filler.

# News digest — <window> (<today's date>)

## AI — releases
- **<Headline>** — <1–2 sentence summary>. [Source](url) · <YYYY-MM-DD>

## AI — practical / how-to
- **<Headline>** — <what it is, why it's useful>. [Source](url) · <YYYY-MM-DD>

## <Your custom category>
- ★ **<Headline>** — <summary>. [Source](url) · <YYYY-MM-DD>
  **What to do:** <one line>

---
**Action shortlist** (★ items only):
1. <action>

Rules:

• Max 5 items per section. More than that, pick the highest-signal ones.
• No item without a working source link. Can't link it → drop it.
• Date every item. Can't verify the date → mark · (date unclear) and demote.
• Lead with ★ items in each section.
• Action shortlist only appears if there's at least one ★ item.

---

What NOT to do

• Don't paraphrase from training data — this is news, time-bound. Every claim needs a recent web source.
• Don't include items older than the window unless asked.
• Don't summarize anything you couldn't actually fetch. If the snippet is too thin, drop it rather than guess.
• Don't editorialize or lecture. State what happened, link the source.

---

Extending the source list

If you say "also check X" during a run, add it to the right bucket for this run and offer to persist it to sources.md for next time.