| name | fetch-news |
| description | Pulls the latest Google News and Hacker News items for every topic and genre in the reader's interests, deduped against every item shown in previous runs. |
Fetch News Skill
Use this skill at the start of every sweep. It queries two public sources —
no account, credentials, or API key required for either — and owns the
cross-run dedupe state:
- Google News RSS (
news.google.com) — a search feed per topic plus a
headline feed per genre (WORLD, NATION, BUSINESS, TECHNOLOGY,
ENTERTAINMENT, SCIENCE, SPORTS, HEALTH)
- Hacker News via the Algolia search API (
hn.algolia.com) — stories and
comments per topic, for the community-buzz angle
- Medium RSS (
medium.com) — long-form posts per tag from
interests.json's medium_tags, for the practitioner-writing angle
Usage
python3 skills/fetch-news/scripts/fetch_news.py
Topics, genres, and the per-topic cap all come from
.agents/workspace/interests.json — change that file, not the script.
The script:
- Reads the interests profile.
- Fetches every topic (both sources) and every genre (Google News). If a
source fails for one topic, the sweep continues and records the error in
source_errors.
- Drops every item whose id is already in
.agents/workspace/seen_items.json,
then adds the new ids to that file (bootstrapping it on the first run).
- Writes the fresh items to
.agents/workspace/new_items.json and prints them
to stdout, with a one-line summary on stderr.
Output format (new_items.json)
{
"topics": ["Gemini API"],
"genres": ["TECHNOLOGY"],
"fetched_at": "2026-07-15T09:00:03Z",
"source_errors": [],
"new_count": 42,
"items": [
{
"id": "gn:CBMiXyz…",
"source": "news",
"outlet": "The Verge",
"title": "…headline…",
"context": null,
"url": "https://news.google.com/rss/articles/…",
"published": "2026-07-15T08:12:44Z"
source is news (Google News), community (Hacker News), or longform
(Medium, with ids prefixed md: and topic: "medium:<tag>"). engagement
is only present for community items (points or comment count). Genre items
carry topic: "genre:<name>".
Instructions for the Agent
- Run this exactly once per sweep. Running it twice in one sweep marks items
as seen without you having curated them.
- Treat
items as the complete candidate pool for the sweep. new_count: 0
means a quiet interval — still write the briefing.
- If
source_errors is non-empty, name the degraded source in the briefing —
the counts only cover the sources that responded.
