News Monitoring

News Monitoring & Digest Delivery

Set up recurring topic/news monitoring that searches the web for recent articles, compiles a headline digest, and delivers it to a messaging channel or document pipeline.

Workflow

Step 1 — Define Topic

Work with the user to define:

• Subject: what to monitor (e.g. PBSA in Hong Kong, AI regulation, self-hosting tools)
• Geography: which region(s) matter (e.g. Hong Kong, Singapore, global)
• Angles: which aspects to track — investment/deals, policy/regulation, supply/pipeline, occupancy/rental, general commentary
• Velocity: how fast-moving the field is — determines acceptable article age (see Step 5)

Step 2 — Create Search Keywords

Build search queries from the topic definition:

• Multi-angle: one query per angle (investment, policy, supply, etc.)
• Multi-language: English keywords + native language keywords for each geography
• Time-bounded: specify recency in the query where possible (e.g. "2026", "latest")
• Keep queries concise — 3-7 terms works best. Overly long queries return thin results.

Step 3 — Select Sources

Identify which sources matter for the topic and geography:

• English-language: international outlets, local English press, industry publications
• Native-language: local press in the geography's primary language(s) — these often break news first
• Industry-specific: trade publications, research houses, REIT/sector-specific sites
• Closed-ecosystem platforms (小红书, 抖音, 微信公众号): content is NOT web-indexed. Be upfront about this limitation — options are manual monitoring or third-party analytics tools.

Record the source list in the topic's reference file (see Reference Files section).

Step 4 — Run Searches

Date grounding (mandatory). Before running any search, ground the model in the real current date. LLMs default to training-cutoff years in queries — this produces stale results. Inject:

Today's date is {current date as "DD Month YYYY"}. When a search query needs a year or refers to "latest"/"current"/"this year", use {current year} or relative wording — never a year inferred from training data.

Primary method: use the harness web_search tool.

web_search(query="Hong Kong PBSA student accommodation investment 2026", limit=10)

The tool uses whatever backend is configured in config.yaml (e.g. SearXNG → DDGS fallback). The skill is backend-agnostic — it does not prescribe which search engine to use. Run one web_search call per query from Step 2.

Advanced option: SearXNG direct curl for news-specific filters.

When you need news-category filtering, language selection, or time bounding that web_search doesn't expose:

# English news, last month
curl -sL "${SEARXNG_URL}/search?q=ENCODED_QUERY&format=json&categories=news&language=en&time_range=month"

# Chinese news, last month
curl -sL "${SEARXNG_URL}/search?q=ENCODED_ZH_QUERY&format=json&categories=news&language=zh-Hant&time_range=month"

Key SearXNG parameters:

• format=json — mandatory (omitting returns 403)
• categories=news — news-only results
• language=zh-Hant / language=en
• time_range=day|week|month|year
• Results include direct source URLs (no proxy resolution needed)

RSS feeds for curated must-track sources that search engines may not surface quickly:

curl -sL "https://example.com/feed/"

Parse <item> blocks for <title>, <pubDate>, <link>, <description>.

Step 5 — Process Results

Go through all search results and:

1. Quality filter — before any ranking, discard low-quality results:
   • Thin content: page is a landing page, aggregator stub, or has no substantive text (<100 words of actual article content)
   • Irrelevant: result doesn't match any defined topic angle despite keyword overlap (e.g. "port" matching "transport" when the topic is about a seaport)
   • Duplicate URLs: same article appearing in multiple search results (keep the first occurrence)
   • Non-article: video-only pages, image boards, login walls with no preview
   • Use word-boundary matching for topic terms to avoid false-relevance from substring matches (e.g. "us" matching "business")
2. Deduplicate — same article appearing across multiple queries or engines appears once (match by title similarity, not just exact URL)
3. Sort by the user's priority order:
   • Recency: most recent first (default)
   • Relevance: how directly it matches the topic angles
   • Source priority: Tier 1 sources rank higher (see topic reference files)
4. Rank — assign each article a priority/impact score based on recency × relevance × source tier
5. Flag stale content — highlight to the user when results are dated relative to the field's velocity:
   • Fast-moving fields (AI, crypto, tech regulation): articles >2-3 months old may be stale — flag them or exclude
   • Slow-moving fields (real estate, infrastructure, policy): articles up to 6 months old may still be relevant
   • Use the velocity defined in Step 1 to calibrate
6. Filter — remove articles outside the topic scope, outside the geography, or behind the recency threshold

Step 6 — Compare With Previous Digests

If this is a recurring digest (cron job), check for overlap:

• Use context_from on the cron job to inject the previous digest output as context (see Cron Setup section)
• Compare new results against the previous digest — match by title similarity, not exact string match
• Remove articles that already appeared in the previous run
• For niche/low-activity topics where overlap is high, consider broadening scope or reducing frequency

Step 7 — Compile Digest

Format the ranked, deduplicated results into a readable digest:

📰 [TOPIC] — Headline Digest
🗓 [Today's date as DD Mon YYYY]

1. [Date] — Source
   **Headline**
   1-2 line summary. Key stat or implication.
   🔗 Direct source URL

2. [Date] — Source
   ...

---
🔑 Key themes:
- Theme 1: brief summary
- Theme 2: brief summary

Formatting rules:

• Always include the publish date on every item
• Headline + 1-2 line summary (not a paragraph analysis)
• Sort by priority/impact (highest first)
• Source links as 🔗 + direct URL to the publisher's site — never Google News proxy URLs or Bing apiclick URLs
• Use 🆕 for articles from the last 7 days
• Max items: 5 by default (configurable per topic). Reduce frequency rather than padding with stale content.
• Key themes summary at the bottom — one-line takeaways of dominant trends

URL rules:

• Every link MUST resolve directly to the article on the publisher's website
• Google News RSS links (news.google.com/rss/articles/CBMi...) are protobuf-encoded proxy URLs — unresolvable by any method. If using Google News RSS for discovery only, resolve the actual source URL via SearXNG title-search or web_search before including.
• Bing News RSS links are apiclick.aspx redirects — extract the real URL from the url= parameter.
• If resolution fails: include the article with ⚠️ Direct link unavailable — search [SOURCE] for: [title]

Step 8 — Delivery

Deliver the digest to the user via one of:

• Chat channel (Telegram, WhatsApp) — most common. Use deliver parameter on cron jobs (e.g. deliver=telegram, deliver=whatsapp:<HANDLE>). WhatsApp requires an explicit target — bare deliver=whatsapp fails.
• Document generation — pass the compiled digest to a document generation skill to produce a formatted PDF, slide deck, or article. The skill should output the digest content in a structured format (markdown) that the downstream skill can consume.

For one-off digests: deliver directly in the current chat. For recurring digests: set up a cron job (see below).

Step 9 — Cron Setup (Reference)

For recurring digests, create a scheduled job using the cronjob tool. This is optional — one-off digests don't need it.

Key parameters:

• schedule: cron expression (e.g. 0 9 1,15 * * for biweekly at 9am, 0 3 */3 * * for every 3 days)
• prompt: self-contained instructions including topic, search queries, source lists, digest format, output language, and output discipline
• deliver: target channel (telegram, whatsapp:<HANDLE>, etc.)
• context_from: list of job IDs whose previous output should be injected as context (for dedup — typically the job's own ID for self-referencing)
• enabled_toolsets: restrict to ["web", "terminal"] to reduce token overhead

Cron prompt template:

You are a [TOPIC] news scout. Find recent articles (within the last [N] months ONLY) and compile a concise digest.

## Date grounding
Today's date is {current date}. When a search query needs a year or refers to "latest"/"current"/"this year", use {current year} or relative wording — never a year inferred from training data.

## Search
Run [N] searches using the web_search tool — one per angle, in both English and [native language]:

1. web_search("[English query 1]", limit=10)
2. web_search("[English query 2]", limit=10)
3. web_search("[Native language query 1]", limit=10)
...

If results need news-specific filtering (categories, language, time_range), use SearXNG directly:
curl -sL "${SEARXNG_URL}/search?q=ENCODED_QUERY&format=json&categories=news&language=[lang]&time_range=month"

If SearXNG is unreachable, the web_search tool will fall back automatically.

## Processing
1. Quality filter — discard thin content (landing pages, stubs, <100 words), irrelevant results (word-boundary match topic terms, not substring), duplicate URLs, and non-article pages
2. Deduplicate across queries (match by title similarity)
3. Filter: only articles from the last [N] months, directly related to [TOPIC]
4. Compare against previous digest (provided as context) — remove articles already covered
5. Sort by recency, then by relevance
6. Pick TOP [5] articles

## Output
📰 [TOPIC] — Headline Digest
🗓 [today's date]

1. **{Headline}**
   📅 {DD Mon YYYY} · {Source}
   {1-2 line summary}
   🔗 {DIRECT article URL}

...up to [5] items

---
🔑 Key themes: {one-line takeaway}

## Rules
- Every 🔗 link MUST be a direct URL to the publisher's site
- Always include publish date
- Use 🆕 for articles from last 7 days
- Do NOT fabricate articles
- OUTPUT DISCIPLINE: Output EXACTLY ONE copy of the digest. The VERY FIRST LINE must be '📰'. The VERY LAST LINE must be the last theme bullet. No thinking, reasoning, planning, or meta-commentary.
**IMPORTANT: Write ALL output in [LANGUAGE].** Preserve original [native language] headlines as-is when the source article is in [native language].

Dedup via context_from: Set context_from: ["<job_id>"] on the cron job (using the job's own ID) so each run receives the previous digest. Add to the prompt: "Review the previous digest output (provided as context). Do NOT include any article that already appeared — match by title similarity, not exact string match."

Frequency tuning:

• High-volume topics: daily or every 2-3 days
• Medium topics: weekly
• Niche/low-activity topics: biweekly (1st and 15th) — allows more content to accumulate, reduces overlap

Source Selection Reference

Hong Kong Sources

English: SCMP, The Standard HK, Hong Kong Business, Mingtiandi, Colliers/JLL/CBRE/Savills research, ResearchAndMarkets

Chinese (Traditional): HK01, 香港經濟日報 HKET, 明報 Ming Pao, 文匯報 Wen Wei Po, 東方日報 on.cc, Now財經, 信報 HKEJ, 香港商報

Industry-specific: PBSA News, Global Student Living, Centurion announcements (SGX-listed PBSA REIT)

Pitfalls

• No dates in digest = rejected. User explicitly asked for publish dates on every item.
• Stale articles = rejected. Filter strictly by the field's velocity — flag or exclude articles past their relevance window.
• Over-summarizing. User wants headline + 1-2 lines, not a paragraph analysis.
• Too many items. User wants max 5 items by default. Reduce frequency rather than padding with stale content.
• Missing native-language sources. Local-language outlets often break news first. Always search in both English and the geography's native language(s).
• Don't promise closed-ecosystem platforms. 小红书, 抖音, 微信公众号 are app-locked. Be upfront.
• Google News RSS links are unresolvable. Never include news.google.com/rss/articles/CBMi... URLs. Resolve to the publisher's direct URL first.
• Output duplication — digest rendered twice. Some models write the full digest during "thinking" then write it again. Fix: add "OUTPUT DISCIPLINE: Output EXACTLY ONE copy." and specify first/last line anchors.
• Output language drift. The skill contains native-language search queries (input). Multilingual models may interpret this as a signal to produce output in that language. Always specify output language explicitly — anchor at both top and bottom of the prompt.
• "Must-include" RSS sources crowded out by discovery results. Use quotas (e.g. 5 discovery + 5 RSS) or guarantee minimum slots per must-include source.
• WhatsApp cron delivery requires explicit target. deliver=whatsapp fails. Use deliver=whatsapp:<HANDLE> or an explicit chat ID.
• delegate_task with web toolset is unreliable for search-heavy tasks. Run searches directly and parse output yourself.
• SearXNG address changed. If the NAS IP changes, update SEARXNG_URL in ~/.hermes/.env and the curl commands in topic reference files.
• DDGS installed in Hermes venv. No PYTHONPATH hacks needed — ddgs CLI and from ddgs import DDGS work directly. Do not use stale paths like /root/.hermes/ddgs_lib or /opt/hermes/.venv/bin/python.

Reference Files

• references/web-search-fallback.md — How the Hermes web search backend dispatch works, current config (SearXNG → DDGS fallback), runtime fallback patch details, and env var verification tips. Load when debugging search issues.
• references/pbsa-hk-search-config.md — PBSA Hong Kong specific: search queries, source tiers, key themes, delivery config. Load when running the PBSA HK monitoring scan.
• references/selfhosting-digest-config.md — Self-hosting & homelab digest: search queries, RSS must-track sources, delivery config. Load when running the Self-Hosting weekly digest.

Evals

evals/routing-fixtures.json holds lightweight contract fixtures — sample request → expected routing (including when a request should go to deep-research or youtube-topic-research instead), required digest output fields, and forbidden output patterns. They are specs, not run against a live model; the repo-root tests/test_routing_fixtures.py validates they stay well-formed and route to real skills.