Generate Daily News Reports

This skill orchestrates a daily curated news report using a Main Agent (scheduling, monitoring, evaluation, decision, aggregation) that dispatches work to a SubAgent execution layer (parallel WebFetch workers plus a headless-browser worker for JS-rendered pages), with smart caching to avoid re-scraping.

It reads two configuration files: sources.json (source configuration, priorities, scrape methods) and cache.json (cached data, historical stats, deduplication fingerprints). Its eight-phase execution: Phase 1 initializes (determine date, read configs, create the NewsReport/ output directory, check for a partial report to append to); Phase 2 dispatches subagents in waves - Wave 1 runs Tier1 sources (e.g. Hacker News, HuggingFace Papers) in parallel, escalating to a Wave 2 of Tier2 sources if under 15 high-quality items, then a Wave 3 browser worker for JS-rendering-required sources (e.g. ProductHunt, Latent Space) if still under 20; Phase 3 defines the structured task format each subagent receives (source URLs, an output schema of source_id/title/summary/key_points/url/keywords/quality_score, and filter constraints keeping cutting-edge tech/deep tech/productivity content while excluding general science/marketing/overly academic/job posts).

Phase 4 has the Main Agent monitor subagent status and success rate per source, retry or skip failed subagents, disable persistently failing sources, and apply a stopping decision (25+ items with 20+ high-quality stops scraping; under 15 continues to the next batch; exhausting all batches under 20 still generates the report favoring quality over quantity). Phase 5 deduplicates (exact URL match, over-80%-similar titles, cross-referencing cache.json history), calibrates scores across subagents by source credibility, and sorts to take the top 20. Phase 6 uses MCP Chrome DevTools (new_page, wait_for, take_snapshot, close_page) for pages that need JS rendering. Phase 7 generates a Markdown report at NewsReport/YYYY-MM-DD-news-report.md with a title/date, statistical summary, 20 templated high-quality items, and generation metadata. Phase 8 updates cache.json with the run's last-run info, per-source stats, processed URL cache, content hashes, and article history.

Its subagent call examples cover using a general-purpose agent (since custom agents require a session restart to be discovered) injected with a worker prompt instructing it to act as a stateless execution unit returning structured JSON with status/data/errors/metadata, versus using a dedicated worker agent type when available. Its output template includes a warning line when the dedicated worker subagent isn't detected, running in a slower serial-execution fallback mode instead.

Its core constraints: quality over quantity (low-quality content never enters the report), early stop at 20 high-quality items, parallel-first execution within a batch, fault tolerance (one source's failure doesn't derail the run), cache reuse to avoid re-scraping, Main-Agent-only decision authority, and graceful fallback awareness. Expected performance: about 2 minutes when Tier1 sources suffice, 3-4 minutes when Tier2 supplementation is needed, and 5-6 minutes when browser-rendered pages are involved. Its error handling table covers subagent timeouts (log and continue), source 403/404s (mark disabled in sources.json), extraction failures (return raw content for the Main Agent to decide), and browser crashes (skip and log). Its compatibility check detects whether a dedicated worker subagent exists at initialization, falling back to sequential Serial Execution Mode with a mandatory degraded-mode warning in the report header if not.

# Daily News Report (YYYY-MM-DD)

> Curated from N sources today, containing 20 high-quality items
> Generation Time: X min | Version: v3.0

---

• Summary: 2-4 lines overview
• Key Points:
  1. Point one
  2. Point two
  3. Point three
• Source: Link
• Keywords: keyword1 keyword2 keyword3
• Score: 5/5

Use this skill to generate a daily curated news report by dispatching parallel scraping subagents across tiered sources, falling back to headless-browser scraping for JS-heavy pages, deduplicating and scoring results, and caching to avoid re-scraping.

Inputs: a target date (or current date) and the sources.json/cache.json configuration files.

Outputs: a Markdown report at NewsReport/YYYY-MM-DD-news-report.md containing the top 20 deduplicated, quality-scored items with summaries, key points, and keywords, plus an updated cache.json.

MCP Chrome DevTools (new_page, wait_for, take_snapshot, close_page) for JS-rendered pages; general-purpose or dedicated worker subagents for parallel scraping; sources like Hacker News, HuggingFace Papers, ProductHunt, and Latent Space.

Users wanting an automated, daily curated tech/productivity news digest generated via parallel subagent scraping with quality filtering and caching.