Agentic news digest

Agentic News Digest 📰

Turn information overload into a curated knowledge stream. This skill enables an agent to act as a news curator for a team -- collecting tips from conversations, researching sources, fact-checking claims, scoring relevance, and delivering formatted digests.

When to Use

• Running a Daily News Digest for a team channel
• Publishing Breaking News about significant industry events
• Announcing new skills, features, or milestones internally
• Creating a Weekly Recap of what happened
• Setting up a Tech Radar to track specific trends
• A team member shares a link and says "put this in the next digest"

Setup

On first activation, create memory/news_digest_config.json:

{
  "audience": "Team name or description",
  "goals": ["What the digest should achieve"],
  "detail_level": "Brief | Balanced | Deep-Dive",
  "sources": {
    "external": ["RSS URLs", "Twitter accounts", "blogs"],
    "internal": ["GitHub org/repos", "internal docs"]
  },
  "channel": "Channel ID for delivery",
  "schedule": "HH:MM Timezone",
  "custom_topics": []
}

---

The Inbox Problem: Why This Section Exists

Agents often receive news tips in conversational sessions (DMs, group chats) but publish digests from isolated cron sessions. These are separate processes with no shared memory.

If you only "remember" a link in conversation, the cron job will never see it.

The inbox is a file-based handoff mechanism that solves this. It is the ONLY reliable way to pass items from conversation → digest pipeline. Not memory files, not MEMORY.md, not "mental notes." The inbox file.

Inbox File: news/inbox.json

{
  "version": 1,
  "items": []
}

Also create news/archive/ for processed items.

Item Schema

Each inbox item looks like this:

{
  "url": "https://example.com/article",
  "title": "Optional title if already known",
  "context": "Why this was shared / what the submitter said about it",
  "draft": null,
  "submitted_by": "Name of the person who sent it",
  "submitted_at": "2026-03-31T10:30:00Z",
  "priority": "normal | high",
  "target_type": "daily | breaking | skill-announcement | milestone"
}

Receiving Tips During Conversation

When a team member shares a link or news tip in a DM or group chat:

1. Read news/inbox.json
2. Append the item with full metadata (URL, context, who sent it, timestamp)
3. Write back to news/inbox.json immediately
4. Confirm to the user that it's queued

Do NOT:

• Write it to a separate queue file that the cron job doesn't check
• Store it only in memory/daily notes and hope it gets picked up
• Say "vorgemerkt" without actually writing to news/inbox.json
• Rely on session context surviving to the next heartbeat

The rule is simple: If it's not in news/inbox.json, it doesn't exist for the digest pipeline.

Co-Authored Drafts

Sometimes a team member doesn't just drop a link -- they collaborate with the agent on the article text in conversation. They iterate, give feedback, adjust wording. The result is a co-authored draft that both sides agreed on.

Two submission modes:

Mode	What happens	Agent's job
Raw input (link, bullets, notes)	Agent writes the article from scratch	Research, summarize, format
Iterated draft (text refined in conversation)	Agent publishes as-is	Format-fix only (platform syntax, typos)

Rules for iterated drafts:

• Store the final agreed text in the draft field of the inbox item
• Set priority: "high" automatically
• The digest pipeline MUST publish the draft verbatim -- no rewriting, no re-summarizing
• Only allowed changes: platform formatting (e.g., Slack *bold* syntax) and typo fixes
• Fact-checking still applies (verify claims/links work), but the text stays as agreed
• Confirm to the submitter: "Gespeichert in der Inbox -- wird genau so gepostet."

Why this matters: When someone invests time iterating on text with you, they chose every word deliberately. Re-summarizing it in the cron job throws away that effort and breaks trust.

---

The Digest Pipeline

Every digest -- daily, weekly, or breaking -- follows these 5 steps in order.

Step 1: Read the Inbox (ALWAYS FIRST)

Before any external aggregation:

1. Read news/inbox.json
2. Separate items into two buckets:
   • Draft items (draft is not null): Auto-include. Skip scoring & summarization.
   • Raw items (draft is null): Feed into the normal pipeline with a +2 relevance bonus.
3. If the inbox is empty, proceed to external aggregation.

This step is not optional. The inbox exists because team members explicitly flagged these items as relevant. Skipping it defeats the purpose of the entire inbox mechanism.

Step 2: Aggregate External Sources

Gather raw candidates from configured sources:

Source Type	Method	Notes
RSS/Atom feeds	web_fetch on feed URLs	Parse for items < 24h old
Twitter/X	RSS proxies or x-reader skill	Check configured accounts
GitHub repos	gh api for commits, PRs, releases	Last 24h for daily, 7d for weekly
Web search	web_search with specific queries	Not broad scans -- targeted topics
Exa AI	exa-search-api skill (if available)	Best for semantic discovery
Internal	Project files, recent team activity	Only company-public information

Target: 15-25 raw candidates for a daily digest, 40-60 for a weekly.

Step 3: Research & Fact-Check

For each candidate (external AND raw inbox items), verify:

Freshness:

• Extract the publication date from the source (not just when you found it)
• Daily: published within last 24h. Weekly: within last 7 days.
• No date available → verify via web_search that the topic is current
• ⚠️ NEVER present old news as new. When in doubt, drop the item.

Source quality:

• Is the source authoritative? (Official blog, verified account, reputable outlet)
• Can the claim be corroborated by a second source?
• For numbers/stats: trace to the original study or announcement
• For product launches: verify on the official product page

Context enrichment:

• What's the backstory? (Follow-up to something the team already knows?)
• Competing perspectives? (Don't present one-sided narratives)
• Actual impact vs. hype?

Drop the item if: No verifiable date, single anonymous source, contradicted by official channels, clickbait with no substance, older than the time window.

Step 4: Score & Select

Score each verified item on a 0-10 scale:

Criterion	Weight	Description
Strategic Fit	3	Directly relevant to team's current projects or strategy
Actionability	2	Team could act on this (adopt tool, adjust approach)
Novelty	2	Genuinely new information, not a rehash
Impact	2	Significant industry shift, major launch
Source Quality	1	Authoritative, primary source

Bonuses:

• Inbox submission: +2 (team member already deemed it relevant)
• Co-authored draft: auto-include (bypasses scoring entirely)

Thresholds:

• ≥ 7: Include (highlight if ≥ 9)
• 4-6: Include if space permits
• < 4: Drop

Item limits: Daily: 5-8 items total. Weekly: 10-15. Quality over quantity -- always.

Step 5: Format, Publish, Archive

1. Select the appropriate template (see below)
2. Apply platform-specific formatting
3. Publish to the configured channel
4. Move processed inbox items to news/archive/inbox_YYYY-MM-DD.json
5. Reset news/inbox.json items to []
6. Update memory/heartbeat-state.json (e.g., last_news_digest: "YYYY-MM-DD")
7. Log in memory/YYYY-MM-DD.md

---

Article Templates

📰 Daily Digest

📰 _Yesterday Daily -- [Weekday], [DD. Month YYYY]_

--

🏠 _Intern bei Yesterday_

• _[Title]_ -- [1-2 sentence summary]. <URL|→ Details>
[If no internal news: "Ruhiger Tag intern -- keine nennenswerten Updates."]

--

🌐 _Tech & AI News_

• _[Title]_ ([Source], [DD.MM.])
  [2-3 sentences: what happened + why it matters for us]
  <URL|→ Quelle>

[Repeat for top 5-7 external items, sorted by score descending]

--

_Kuratiert von [Agent Name] 🐾 | Quellen: [X] geprüft, [Y] relevant_

Rules:

• Internal section ALWAYS comes first, NEVER skip it
• Each item: source name, date, working link
• Max 8 items total -- this is a digest, not a dump
• Inbox items with drafts: insert the draft text verbatim in the appropriate section

⚡ Breaking News

For single high-impact events that can't wait (score ≥ 9, at least 2 sources).

⚡ *Breaking: [Headline]*

[3-5 sentences: what happened with concrete details]

*Warum das wichtig ist:*
[2-3 sentences connecting to team's work]

*Quellen:*
• <URL1|[Source 1]> ([Date])
• <URL2|[Source 2]> ([Date])

_[Agent Name] 🐾 | [Timestamp]_

Max 1 breaking news per day. Multiple breaking items → bundle into urgent daily digest.

🐾 Skill / Feature Announcement

🐾 *Neuer Skill: [Name] -- [One-line description]*

[2-3 sentences: what it does, what problem it solves]

*Was steckt drin:*
• [Capability 1] -- [brief description]
• [Capability 2] -- [brief description]

*Warum das wichtig ist:*
[1-2 sentences: concrete benefit for the team]

→ <URL|Zum Skill>

Keep under 200 words. Focus on value, not implementation.

🏆 Milestone Update

🏆 *Milestone: [What was achieved]*

[2-3 sentences: significance of the achievement]

*Key Facts:*
• [Measurable detail]
• [Measurable detail]
• [What's next]

_[Agent Name] 🐾_

Only for genuine milestones, not routine progress.

📊 Weekly Recap

📊 _Yesterday Weekly -- KW [XX] ([DD.MM.] - [DD.MM.YYYY])_

--

🏠 _Diese Woche intern_

• _[Project/Theme]_ -- [Week summary]. <URL|→ Details>
[Group by project, not by day]

--

🌐 _Die wichtigsten Tech-News der Woche_

1. _[Title]_ ([Source], [DD.MM.])
   [Summary + relevance]
   <URL|→ Quelle>

[Top 5-8 items of the entire week, numbered by importance]

--

📈 _Zahlen der Woche_ (optional)
• [Interesting stat]
• [Interesting stat]

--

_KW [XX] Recap von [Agent Name] 🐾 | [X] Quellen ausgewertet_

Don't concatenate 5 daily digests -- curate the week's highlights.

---

Platform Formatting Reference

Platform	Bold	Italic	Links	Tables	Headers
Slack	*bold*	_italic_	<URL|Text>	❌ No	❌ No
Discord	**bold**	*italic*	[Text](URL) or suppress with <URL>	❌ No	✅ #
WhatsApp	*bold*	_italic_	Plain URLs	❌ No	❌ Use CAPS

---

Quality Checklist (Run Before Every Publish)

• news/inbox.json was read and processed (inbox items included or archived)
• Every item has a verifiable publication date
• Every item has a working source link
• No item is older than the digest's time window
• Each item has a "why this matters for us" line
• Internal section is present (even if "quiet day")
• Platform formatting is correct for the target channel
• No duplicate topics from yesterday's digest
• Co-authored drafts are published verbatim (not re-summarized)
• Total length stays reasonable (daily ≤ 500 words, weekly ≤ 800 words)
• Language matches channel config (e.g., German for Yesterday)
• Numeric claims cite a source

---

Scheduler Integration

With Heartbeat: Add to your HEARTBEAT.md or equivalent:

1. Check heartbeat-state.json → if last_news_digest == today, SKIP
2. Read news/inbox.json FIRST
3. Run aggregation pipeline
4. Publish, archive inbox, update state

With Cron: Create a dedicated cron job that runs the digest pipeline at the configured schedule. The cron session has NO access to main session memory -- the inbox file is the bridge.

Key principle: The inbox is the ONLY handoff between conversational sessions and digest runs. Everything else (memory files, session context, "mental notes") is unreliable across session boundaries.

---

Maintenance

• Archive processed inbox items: news/archive/inbox_YYYY-MM-DD.json
• Archive daily raw items: memory/archive/news_YYYY-MM-DD.md
• Update memory/heartbeat-state.json after each publish
• Review relevance scoring weights quarterly with the team
• Rotate stale RSS sources quarterly
• Monitor inbox for items that were submitted but never published (indicates pipeline bug)

---

"Signal over Noise. Wisdom over Data." 📡