news-collection

News Collection

Gather news items on a topic, filter junk cheaply, verify freshness, and emit qualified items under an evidence contract. Pipeline-shaped: four phases, a gate between each. Content pipelines consume the JSON artifact; pair with fact-check to verify what this skill qualifies.

Instructions

Phase 1: COLLECT

Goal: gather candidate items from available sources (feeds, search results, provided fixtures). Every item carries the five-fact evidence contract: title, url, outlet, author, published_at.

Rule (verbatim from the design): publish times are extracted (article metadata), never guessed; missing timestamp → recorded as unknown, confidence lowered and disclosed. A guessed timestamp poisons every downstream freshness verdict; an honest "published_at": null keeps the item usable with known uncertainty.

Record each item as one JSON object per the item schema below. Fill evidence_notes with where each fact came from (meta tag, byline, JSON-LD, sitemap).

Distinguish outcomes: zero items because sources were unreachable is a collection failure (report it, stop); zero items from reachable sources is a valid empty feed (deliver an empty artifact with counts of zero).

Gate: every collected item has all five fields present — value or explicit null with a confidence downgrade and a disclosure note. Items with silent gaps stay in COLLECT until the gap is recorded.

Phase 2: COARSE FILTER

Goal: cheap, high-recall pass over collected items. Three verdicts: keep / monitor_only / reject, each with a reason code (see Coarse Filter below).

Rule (verbatim from the design): high-magnitude stories are downgraded to monitor_only at most, never rejected — a false keep is cheap, a silent drop is expensive. A keep costs one extra freshness check; a wrongly dropped major story costs the whole pipeline its value.

This phase runs on a cheap model when dispatched -- it needs recall, not judgment depth. Give the dispatched model: topic, verdict/code tables, asymmetry rule, and items' five facts plus a text excerpt.