AI Newsletter Daily

Generate a concise daily AI newsletter for a Chinese audience from fresh web sources.

Use this skill only when the request is about current AI/ML news, releases, research, funding, product launches, model updates, regulation, benchmarks, or practitioner-relevant developments.

Do not use this skill for:

• Evergreen explainers.
• Non-AI topics.
• Long-form research that is not intended to become a curated newsletter.

Inputs

Expected inputs, with defaults if missing:

• target_news_count = 20
• search_query = "latest AI news today"
• search_time_window_days = 2
• max_search_results = 60
• min_articles_required = 10
• include_domains = []
• exclude_domains = ["youtube.com", "reddit.com", "facebook.com", "x.com", "twitter.com"]
• summary_model = "host-default"
• max_scrape_retries = 2

Rules:

• Clamp target_news_count to 1..50.
• Clamp search_time_window_days to 1..14.
• Clamp max_search_results to 20..120.
• Clamp min_articles_required to 1..50.
• Clamp max_scrape_retries to 0..5.
• If min_articles_required > target_news_count, set min_articles_required = target_news_count.

Batch policy

Use a two-stage batch limit:

• Search batch: collect up to max_search_results candidates from search.
• Scrape batch: keep the top target_news_count * 2 ranked candidates for fetch and summary attempts.
• Final batch: return only the top target_news_count verified items.

Do not summarize every search result. Over-collect, filter, verify, then reduce to the final batch.

Required outputs

Return all of the following:

1. newsletter_items as a list of objects.
2. markdown_newsletter as a string.
3. json_newsletter as an object.

Each newsletter item must include:

• title
• url
• domain
• published_at
• summary
• relevance_score
• source_query

Use "unknown" for published_at when no date is available.

Deterministic workflow

1. Resolve inputs.

   • Apply defaults and bounds.
   • Initialize warnings = [].
   • Initialize seen_canonical_urls = set().
   • Initialize processed_urls = set().

2. Search.

   • Run web_search with search_query.
   • If there are no usable results, retry once with:
     • "{search_query} generative AI LLM model open source enterprise"
   • If there are still no usable results, fail with a clear message.

3. Normalize and filter.

   • Keep only results with non-empty title and URL.
   • Canonicalize URLs by lowering the host, removing tracking parameters when possible, and normalizing safe trailing slashes.
   • Drop duplicates by canonical URL.
   • Apply include_domains and exclude_domains.
   • Prefer results likely within search_time_window_days.
   • Keep unknown dates, but score them lower.

4. Rank.

   • Score each candidate from 0 to 100:
     • AI-topic relevance: 0..50.
     • Freshness: 0..30.
     • Title/snippet clarity: 0..20.
   • Sort by:
     • relevance_score descending
     • published_at descending, unknown last
     • url ascending
   • Keep the top target_news_count * 2 candidates.

5. Verify and summarize.

   • Process candidates in ranked order until target_news_count verified items are collected.
   • Skip candidates whose canonical URL is already in processed_urls.
   • Attempt web_fetch up to max_scrape_retries + 1 times.
   • If fetch fails, add a warning with the URL and reason, then continue.
   • Cross-check search result vs fetched page using title similarity, domain consistency, topic alignment, and published date when available.
   • If the page appears materially inconsistent, skip it and warn.
   • Summarize each accepted article in one short plain-text paragraph, max about 80 words, focused on why it matters to AI practitioners.

6. Minimum quality gate.

   • If collected items are fewer than min_articles_required, run one fallback search with:
     • "AI news today machine learning model release funding research"
   • Process only new candidates not already seen or processed.
   • Repeat filtering, ranking, verification, and summarization.

7. Final integrity check.

   • Ensure every final item has non-empty title, url, domain, summary, source_query, and numeric relevance_score.
   • Ensure each URL appears once.
   • Ensure markdown_newsletter and json_newsletter match in item count.
   • Remove and warn on any invalid item.

8. Finalize.

   • Sort by relevance_score descending, then published_at descending.
   • Truncate to target_news_count.
   • Render markdown_newsletter.
   • Assemble json_newsletter.
   • Apply the language output rule below.
   • Return all outputs.

Language output

• Translate the final markdown_newsletter body and each article summary in newsletter_items into Simplified Chinese.
• Keep title, url, domain, published_at, relevance_score, and source_query unchanged.
• If a source title is already in Chinese, preserve it as-is.
• Do not add extra commentary outside the newsletter content.

Verification

Accept items only if:

• URL is valid and canonicalized.
• Search result and fetched page broadly match.
• Topic is actually AI/news relevant.
• Published date is present or safely unknown.
• Fetched content is not malformed or off-topic.

Record warnings for failed URLs, short reasons, and whether fallback search was used.

Output Format

markdown_newsletter:

• H1 title with date.
• One H2 per article.
• One short summary paragraph per article.
• One source link per article.

json_newsletter:

• date
• query
• count
• articles
• warnings

Safety rules

• Use only sanctioned tools.
• Do not request API keys from the user.
• Do not expose secrets.
• Do not include copyrighted full article text.
• Keep summaries neutral, concise, and factual.
• Preserve deterministic behavior wherever tool outputs allow.