AGENT SKILL

news-rss-watch

Polls a seeded, deterministic RSS/Atom feed registry covering US news, China news, Singapore news, global market news, important/world news, and arXiv research feeds — with per-feed watermark dedup so each run surfaces ONLY new items. Use whenever the user asks to "check the news", "poll feeds", "what's new since yesterday", "morning news digest", "watch this feed", "seed/track RSS feeds", wants a comprehensive US+China+Singapore news outlook, or wants new arXiv papers on a finance/ML topic watched. Triggers on phrases like "news RSS", "add this feed", "what's new in the feeds", "daily news brief", "track arxiv papers on X". This is the deterministic data layer — pair with `daily-market-watch` when the user wants a synthesized, source-cited market report rather than raw new items.

License 许可
MIT
最近更新
2026-09-13
Stars
★ 0
Forks
2

01 前言 Front Matter

name
news-rss-watch
description
Polls a seeded, deterministic RSS/Atom feed registry covering US news, China news, Singapore news, global market news, important/world news, and arXiv research feeds — with per-feed watermark dedup so each run surfaces ONLY new items. Use whenever the user asks to "check the news", "poll feeds", "what's new since yesterday", "morning news digest", "watch this feed", "seed/track RSS feeds", wants a comprehensive US+China+Singapore news outlook, or wants new arXiv papers on a finance/ML topic watched. Triggers on phrases like "news RSS", "add this feed", "what's new in the feeds", "daily news brief", "track arxiv papers on X". This is the deterministic data layer — pair with `daily-market-watch` when the user wants a synthesized, source-cited market report rather than raw new items.

02 资源文件 8

03 SKILL.md 正文

news-rss-watch

Deterministic, seedable news feed watcher. Polls a config-driven feed registry (feeds.json) with watermark dedup — each run emits only items published since the last run, as strict JSON. No LLM interpretation in the fetch loop: the agent runs the script and renders the output.

Works with any agent

This skill is agent-agnostic by construction:

  • Pure CLI + JSON contract. The scripts take flags, print strict JSON/markdown to stdout, and exit with a status code. No agent framework APIs, no tool hooks, no environment variables beyond NEWS_WATCHER_STATE_DIR / NEWS_WATCHER_DIGEST_DIR (both optional).
  • Only needs a shell. Any agent with a terminal tool (Claude Code, Codex, Cursor, OpenCode, Gemini CLI, plain cron, a human) can drive it — the SKILL.md is instructions for whatever agent is running it.
  • Stdlib-only Python. No pip installs, no virtualenv, no network calls beyond fetching the feeds themselves.
  • Deterministic output. Identical inputs → identical JSON, so any agent parses the same contract. The skill never depends on the agent's memory or prior turns.

When to use

  • "Check the news" / "what's new since yesterday" / "poll the feeds"
  • "Daily news digest" for US + China + Singapore + markets + world
  • "Add this feed / track this website" — deterministic --action add
  • "What feeds are you tracking?" — --action list
  • "Watch arXiv for new papers on <topic>" — scripts/watch_arxiv.py
  • Any cron / scheduled briefing job: run the script, parse JSON, stay silent when the item count is 0.

Mental model

  1. feeds.json — the seed registry. Pre-seeded with live-verified feeds (checked 2026-08-08). Agent never edits feeds by hand; it uses --action add / --action list so the registry stays deterministic.
  2. scripts/watch_feeds.py — batch poller. One command, all feeds, one JSON result. Per-feed isolation: a broken feed is reported in errors, never crashes the batch.
  3. Watermark state — ~/.finance-skills/watcher-state/newsrss_<id>.json (override with NEWS_WATCHER_STATE_DIR). First run = baseline: records IDs, emits nothing. This prevents flooding the agent's context with old articles on the first poll.
  4. scripts/watch_arxiv.py — same pattern against the arXiv Atom API (one-shot arXiv search turned into a watcher).

Feed registry (seeded categories)

Category Focus Seed feeds
us US news + policy + regulators NPR, CNBC, WSJ Markets, SEC press, Fed press, Google News US
china China news (first-party + HK) SCMP news/business, CGTN, Xinhua English, Google News CN (zh)
singapore SG local news + markets CNA, Straits Times, Business Times SG, Google News SG, SG business topic
markets Global market news MarketWatch, Yahoo Finance, FT, BBC Business, Investing.com, CoinDesk, Cointelegraph, Google Business
important World / headline news BBC World, Guardian World/Business, Google News World
research arXiv finance + AI q-fin.ST/EC/TR/PM/MF/RM/GN, econ.GN, cs.AI (weekday-only updates)

Region-source rule (mirrors ADR-002): China category uses Chinese first-party / HK sources (Xinhua, CGTN, SCMP) — not US outlets covering China. Singapore uses local outlets only. See references/feed-registry.md for verification notes and feeds that were tested and rejected (dead, paywalled, rate-limited).

Workflow

News poll run:
- [ ] 1) Pick scope: all categories, or --categories us,china,...
- [ ] 2) Run: python scripts/watch_feeds.py [--categories ...] [--max N] [--budget N]
- [ ] 3) Parse the JSON. count == 0 → reply "No new updates." Do NOT fabricate news.
- [ ] 4) Render stories grouped by category, with title + URL + score.
- [ ] 5) On user request, synthesize: use daily-market-watch for the full report.

Polling (the default action)

python scripts/watch_feeds.py
# → JSON: {"count": N, "first_run": bool, "results": [{category, clusters: [...]}], "errors": [...]}

Clustering is ON by default (anti-noise): the same story picked up by N feeds becomes ONE story. Each cluster:

{
  "cluster_id": 0, "score": 3.5, "matched_keywords": ["fed", "cpi"],
  "category": "markets", "feed_id": "marketwatch_top",
  "feed_name": "MarketWatch Top Stories",
  "cluster_size": 2, "other_sources": ["Google News Business"],
  "items": [{"id", "title", "url", "summary", "published"}]
}
  • score = source weight (3 = regulator/first-party, 2 = major outlet, 1 = aggregator) + 0.5 per importance keyword hit (fed, cpi, tariff, mas, pboc, war, earnings, …). Sort by score when rendering.
  • Attribution: the highest-weight source wins the category + link; other_sources lists the rest. A China story covered by Xinhua + BBC stays in china, not important.
  • Budget: --budget N caps stories per category (default 8, 0 = unlimited).

Behavior flags:

  • First ever run records a baseline and returns count: 0 with first_run: true — tell the user the watcher is primed, then re-run (or use --fresh for a full digest).
  • --categories us,markets — scope to one or more categories.
  • --fresh — emit everything, ignoring watermarks (still records state so the next normal poll stays quiet).
  • --max 15 — per-feed cap (Google News feeds default to 15).
  • --with-summary — include a ≤500-char summary snippet per item.
  • --no-cluster — flat per-feed view (debugging single feeds; no dedup).
  • --format markdown — machine-shaped markdown for cron delivery.

Daily digest (cron-friendly)

python scripts/watch_feeds.py --action digest --max 5 --budget 5
# → writes ~/.finance-skills/digests/news-digest-YYYY-MM-DD.md (override
#   dir with NEWS_WATCHER_DIGEST_DIR), prints {"digest": path, "stories": N}

Fresh full poll → cluster → per-category markdown file. Zero LLM needed: point a nightly cron at it and ship the file to your inbox/channel.

Adding a feed (deterministic, no file editing)

python scripts/watch_feeds.py --action add --name "my_feed" --url "https://example.com/rss" --category us

Validation: --category must be one of us, china, singapore, markets, important, research. Duplicate url/id → hard error, no mutation.

Search-as-feed (gap filler)

# Any topic/outlet Google News can cover — e.g. MAS has no RSS:
python scripts/watch_feeds.py --action add-search \
  --query "MAS Singapore monetary policy" --category singapore \
  --hl en-SG --gl SG --ceid SG:en

Builds a Google News search RSS feed and adds it like any other feed. Region-scoped via --hl/--gl/--ceid (defaults en-US/US/US:en).

Checking / resetting

python scripts/watch_feeds.py --action list                 # show registry
python scripts/watch_feeds.py --action check                # verify every feed URL + XML
python scripts/watch_feeds.py --action reset --feed-id cna_main   # clear one watermark
python scripts/watch_feeds.py --action reset                # clear all watermarks

Watching arXiv topics (research category)

python scripts/watch_arxiv.py --query "cat:q-fin.ST AND agent" --name qfin_agents
python scripts/watch_arxiv.py --query "ti:LLM trading" --max 5
python scripts/watch_arxiv.py --query "all:deep learning volatility" --fresh

Each run emits only papers new since the last run (watermark keyed on arXiv id). Exactly 1 API request per run — safe for cron. Supports the full arXiv query syntax (all:, ti:, abs:, au:, cat:, AND/OR).

Guardrails

  • Never fabricate news. If count is 0, respond with no-new-updates. If a feed errored, report it from errors — never substitute another feed's items.
  • Never edit feeds.json by hand. Use --action add / add-search. This keeps state deterministic and auditable.
  • Never rename a feed id mid-stream. The watermark is keyed on the id; renaming resets dedup and floods context with old items.
  • First-run semantics are a feature. Baseline recording prevents replay. Use --fresh only when the user explicitly wants a full digest.
  • Clustering is deterministic, not editorial. Two stories are merged only when title tokens overlap ≥60% after stopword/outlet stripping. Do not hand-merge or hand-split clusters when rendering; show what the script emits. If clustering looks wrong (over-merge of distinct stories), report it — the threshold lives in the script.
  • Score is a sort key, not a verdict. Higher score = more keywords + more authoritative source. Render in score order but never drop a story the user asked about just because it scored low.
  • Per-feed failure ≠ batch failure. One dead feed goes to errors; the rest still poll. Report errors to the user, don't retry endlessly (respect HTTP 429 by backing off — the next scheduled poll retries).
  • High-churn feeds are capped. Google News feeds emit dozens of items; max_per_poll in feeds.json caps them at 15, --budget caps stories per category.
  • Don't re-order or re-parse feeds yourself. The script is the single source of truth; if a feed breaks, run --action check.

Tests

python3 -m unittest discover -s skills/news-rss-watch/tests

27 offline tests: RSS/Atom parsing (incl. malformed-CDATA repair), title normalization, clustering, importance scoring, budget caps, watermark baseline/dedup/corruption, feed-add validation, search-URL building, digest writing, arXiv Atom parsing. No network needed.

Companion skills

  • daily-market-watch — when the user wants a synthesized, source-cited market report (6 sections, region zoom) instead of raw new items: poll first with this skill, then hand the new items to daily-market-watch as the news layer.
  • smart-money-tracker — when a news item mentions a 13F/STOCK Act filing, cross-check with the tracker.

References

  • references/feed-registry.md — how feeds were verified, what was rejected and why, notes on flaky feeds (arXiv weekend silence, SEC User-Agent requirement, Reddit rate limits, MAS has no working RSS).
  • feeds.json — the live seed registry (edit via --action add only).

Example

  • examples/sample-report.md — a real clustered poll run (all 5 news categories, scores, dedup evidence). Golden output for the eval set; use it as the shape reference when rendering.
  • examples/sample-digest.md — a real nightly digest file (--action digest). Use as the shape reference for cron delivery.

04 同仓库更多 Skills

kangcodex
kangcodex/finance-skills

thematic-stock-picker

High-conviction thematic 5-year stock picker inspired by @crypto_condom — generates 5 distinct 2026–2031 policy-anchored themes and a screened basket of 5–8 US-listed/ADR small/mid-caps per theme. Use whenever the user asks for thematic stock picks, 5-year investment thesis, policy-driven stock ideas, small/mid-cap value screen, contrarian long-only watchlist, "what stocks will benefit from [policy/tech trend]", or any request to produce a multi-year thematic equity basket with screening workings and source citations. Triggers on phrases like "thematic stock picks", "5-year stock thesis", "policy tailwind stocks", "screen me small/mid-cap names", "high conviction basket", "what should I own for the next 5 years", "give me 5 thematic predictions", "stocks for [AI / defense / energy / reshoring / biotech / etc.] theme", or any "5 thematic predictions + baskets" prompt. Always shows the screening workings (universe → filters → survivors) and cites sources for every pick. Do NOT use for: short-term trade ideas, single-ticker deep-dive valuation, options strategies, market commentary without picks, or non-equity assets.

★ 0
kangcodex
kangcodex/finance-skills

test-skill

A test skill for unit testing

★ 0
kangcodex
kangcodex/finance-skills

smart-money-tracker

Track SEC 13F institutional holdings + congressional STOCK Act trades + House Clerk disclosures, produce convergence analysis between whales, politicians, and retail flows. Use whenever the user asks about hedge fund 13F filings, Berkshire/Bridgewater/Pershing Square/Soros/ARK portfolio changes, what Michael Burry or any other major investor is buying, politician stock trades (Pelosi, Tuberville, Cruz, etc.), STOCK Act disclosures, congressional insider trading, "smart money" flows, institutional positioning, whale conviction, cross-signal analysis between 13F and congressional trades, or wants a "smart money report" / "follow the whales" / "what are insiders doing" report. Triggers on phrases like "13F", "whale trades", "politician trades", "congress trading", "Pelosi bought", "smart money", "institutional holdings", "hedge fund positions", "Soros fund", "Berkshire portfolio", "what are insiders buying", "congressional insider", "follow the smart money", "rerun smart money report". Do NOT use for: single-ticker technical analysis, trade execution, options strategies, market commentary without picks, or non-equity flow analysis (use daily-market-watch or thematic-stock-picker instead).

★ 0
kangcodex
kangcodex/finance-skills

sg-financial-advisor

Singapore-licensed Financial Advisor (FA) skill that diagnoses a user's insurance + CPF portfolio, computes the LIA gap against national protection layers, prioritises an action checklist, and flags surrender, ILP, and Premium-Overhead risks under MAS FAA 2001, LIA, and IBF guidelines. Use whenever the user asks for a Singapore insurance/financial-plan review, a "should I surrender" verdict on a Whole Life / Endowment / ILP / Legacy policy, a portfolio gap analysis (Death, TPD, CI, ECI, PA, Hospitalisation), a CPF OA/SA allocation or cashflow analysis, an Integrated Shield / CareShield / HPS review, or a "is my plan good or not" diagnosis. Triggers on phrases like "review my insurance portfolio Singapore", "should I surrender my whole life policy", "is my ILP worth keeping", "GE GreatLink FlexiPlan review", "LIA gap analysis", "MediShield vs Integrated Shield", "CareShield opt-in/out", "CPF OA vs SA allocation", "FAA balanced scorecard", "RES5 surrender warning", "premium overhead inefficiency", "29 year old financial plan SG", or any "is my Singapore financial/insurance plan good" prompt. Always show the gap math, the surrender warning, the action checklist, and the disclaimer. Do NOT use for: tax-only questions, US/EU/ HK/AU-specific products, unit trust/equity-only reviews (use thematic-stock-picker or smart-money-tracker), or any recommendation that constitutes regulated advice outside an FA-rep's licence.

★ 0
kangcodex
kangcodex/finance-skills

daily-market-watch

Generate a "Daily Market Watch" / global finance news report covering the last 7 days for a diversified portfolio — stocks, ETFs, cryptocurrencies, indices, bonds, commodities, Fed signals. Use whenever the user asks for a market overview, daily/weekly market briefing, what is moving markets, global finance news, region-specific market report (e.g. "zoom into China", "US market today", "Japan markets this week"), Fed/rate-cut outlook, value/momentum stock picks, crypto market recap, or anything that sounds like a "morning note" / "market wrap" from a sell-side desk. Triggers on phrases like "daily market watch", "market overview", "weekly market report", "global finance news", "what's happening in markets", "give me a market briefing", "finance report for [region]", "morning market note", "end-of-day report", or any request to summarize market-moving news with sources. Do NOT use for: single-ticker deep dives, trade execution, tax advice, or pure technical chart analysis.

★ 0