openmark-newsletter-roundup

name	openmark-newsletter-roundup
description	Compose a categorical news-roundup newsletter from Ahmad's OpenMark saves. Use when Ahmad wants a scannable "what happened in AI / tools / research / discourse" recap rather than a single-thesis essay. Trigger phrases include "roundup", "news roundup", "this week in AI", "weekly news", "what dropped this week", "category recap", "/newsletter-roundup". Output is bucketed by topic with 3–5 items per bucket, short blurbs, source-tagged for at-a-glance scanning.
metadata	{"type":"composition"}

OpenMark — Newsletter Roundup

Categorical news-roundup format. NOT thesis-driven — this is a scannable recap of what happened, organized by topic bucket. Reader skims headers, dives into 1-2 items.

When to use vs other newsletter skills

User intent	Skill
"What's happening in AI this week"	this one (roundup)
"Make me a newsletter on context engineering"	openmark-newsletter (analytical)
"Write me a thoughtful essay on X"	openmark-newsletter-essay
"Compare A vs B"	openmark-newsletter-comparison
"I want a LinkedIn-style post"	openmark-newsletter-thread

Tool sequence

1. Get time window. Default last 7 days. Ahmad can override.
2. Pull recent fresh: find_recent(days=N, n=30) — empty query.
3. Cross-source angle: search_semantic(query="<broad area>", n=20) (whatever broad area Ahmad named, e.g. "AI tooling, models, research").
4. Source-specific gems (parallel): search_linkedin(query=..., n=10) for hot takes, search_youtube(query=..., n=10) for video drops.
5. Cluster. Bucket the deduped hits into 3–6 topical groups:
   • "Models & research" — papers, model releases
   • "Tooling & open source" — github repos, libraries, frameworks
   • "Industry & business" — funding, partnerships, strategy
   • "Discourse & takes" — LinkedIn posts, opinions
   • "Tutorials & courses" — YouTube videos, blog tutorials
   • "Off-topic but interesting" — wildcards worth a click

Don't force all 6 — only use buckets that have ≥2 items.

Output format — NON-NEGOTIABLE

# {Week of YYYY-MM-DD} — {Punchy 4-word headline}

_Quick recap of {N} items I saved {window}. {1-sentence pulse on the week.}_

## {Bucket 1 name}  ({count})

1. **{Title}** — [{domain}]({URL})

   {1-sentence so-what. Why this matters.}

2. **{Title}** — [{domain}]({URL})

   {1-sentence so-what.}

(repeat 3–5 items per bucket)

## {Bucket 2 name}  ({count})

(same pattern)

## Sources cited

1. [{title}]({URL})
2. [{title}]({URL})
...

_{N} items · {hours} of content · feed pulled {window}_

Formatting rules — strictly enforced

• One blank line between every numbered item. Otherwise the chat collapses them into one wrapped paragraph.
• Each item: title in bold, then —, then a short clickable [{domain}]({URL}) link (NOT the full URL pasted as text), then a blank line, then a single-sentence so-what indented or as its own line.
• Bucket headers are ## — never ### or bold-only.
• The trailing ## Sources cited is mandatory so the chat auto-saves the draft.

Voice

Tight, declarative, no hedging. Each so-what is ONE sentence. No "this could be useful," no "I found this interesting." Examples:

Generic	Roundup voice
"This is an interesting paper about RAG."	"Argues every production RAG should use cross-encoder rerank. Has numbers."
"A new library for agents was released."	"Drop-in replacement for LangChain's create_agent. 40% fewer LOC, same hooks."
"Worth checking out."	"Skip unless you're shipping agents this week."

What NOT to do

• Don't editorialize ("the week was interesting") — let the items speak.
• Don't pick more than 5 items per bucket. If a bucket has 10, split into two sub-buckets or drop the bottom 5.
• Don't include URLs you didn't see in a tool result.
• Don't summarize at the end. The buckets ARE the summary.