News Collection Agent Skill

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 schema in

references/evidence-contract.md. 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 from

references/coarse-filter.md.

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. See the dispatch note in references/coarse-filter.md.

Gate: every item has exactly one verdict and one reason code. reject

verdicts on items that look high-magnitude get re-checked once before the

phase closes.

Phase 3: FRESHNESS CHECK

Goal: for each keep and monitor_only item, establish when the story

first became public and whether this page is the original coverage. Methods in

references/freshness-forensics.md:

• First-public time vs page date — a page can carry today's date on old news.
• Syndication and aggregator reposts detected; select the canonical coverage.
• Same-story vs new-development rubric — consolidate duplicates, keep genuine

developments.

Rule (verbatim from the design): two independent sources or verdict "unclear".

Conservative default: unclear over guessed. An "unclear" verdict is

recoverable downstream; a confidently wrong "fresh" verdict ships stale news.

Gate: every surviving item carries freshness: fresh | stale | unclear, a

first_public_estimate (or null), and the count of sources backing the

verdict. Duplicate clusters are consolidated to one canonical item with

duplicates_of links.

Phase 4: DELIVER

Goal: emit qualified items as a structured JSON artifact (schema:

references/evidence-contract.md) plus a summary table. Every verdict state

appears in the artifact — monitor_only, unclear, and reject items ship

with their verdicts rather than vanishing, so consumers see the full triage.

Gate (deterministic phase checkpoint — emit this table before delivering the

artifact; delivery without it is incomplete):

| Verdict | Count |

|---------|-------|

| keep | n |

| monitor_only | n |

| reject | n |

| unclear (freshness) | n |

| duplicates consolidated | n |

The counts make silent drops visible: collected total must equal

keep + monitor_only + reject. If it does not, return to the phase that lost

items.

Error Handling

No items collected

• Cause: sources unreachable, or feed genuinely empty.
• Solution: unreachable sources → report collection failure and stop;

reachable but empty → deliver an empty artifact with zero counts. The two

outcomes stay distinct so an outage is not read as a quiet news day.

No timestamp found anywhere for an item

• Cause: page has no metadata, byline date, or sitemap entry.
• Solution: set published_at: null, confidence: low, disclose in

evidence_notes; freshness verdict for that item is unclear.

Two sources disagree on first-public time

• Cause: syndication chain or republished update.
• Solution: apply canonical-coverage selection

(references/freshness-forensics.md); if still split, verdict unclear.

Item count mismatch at DELIVER

• Cause: an item was dropped without a verdict.
• Solution: diff item ids against the COLLECT record; assign the missing item

a verdict with reason code.

Reference Loading Table

| Signal | Load These Files | Why |

|---|---|---|

| Recording items, JSON artifact schema, confidence fields | evidence-contract.md | Five-fact contract and artifact schema |

| Assigning verdicts, reason codes, cheap-model dispatch | coarse-filter.md | Verdict definitions and dispatch note |

| Dating a story, syndication, duplicates, canonical pick | freshness-forensics.md | Forensic methods and rubrics |