Article Intelligence

Required Output

Always create a dedicated static HTML reading workspace for each article-intelligence run unless the user explicitly asks for chat-only output.

Save the HTML in the current task's user-facing outputs/ directory when one is provided by the environment. Use a timestamped filename such as:

article-intelligence-YYYY-MM-DD-HHMM.html

The HTML report must be the primary deliverable. The chat response should briefly summarize what was generated and link to the HTML file.

HTML Report Requirements

The report must be an interactive static reading workspace, not a long stacked document. It must include:

• A fixed top report header with date range, source set, generation time, and match count.
• A left article rail with searchable article buttons. Clicking an article must update the reading pane without requiring page reload.
• A central reading pane focused on one selected article at a time.
• A right-side inspector on desktop for trend judgment and current-article metadata. Hide or collapse it on smaller screens.
• For every article:
• Original source link.
• Article metadata.
• A prominent original-content panel near the top of the selected article view. Show source URL, visible publication notes, and a structured original-content map or short excerpts. Do not hide this at the bottom of the page.
• Deep analysis: thesis, key claims, evidence, assumptions, missing context, and practical implications.
• Clearly labeled personal take.
• Original article content notes or compliant excerpts. Respect copyright limits: do not reproduce paywalled or excessive copyrighted text; prefer concise excerpts, section maps, and paraphrased notes.
• A final cross-article trend judgment when there is more than one article.

Use scripts/build_article_report.py to turn a structured JSON analysis file into the HTML report.

If the requested date range has no matching articles, generate an empty-result HTML report for that exact range. Do not add older articles, fallback recommendations, or nearby posts unless the user explicitly asks for them.

Core Workflow

1. Resolve the requested time window.
• If the user says "today", use the current local date from the environment.
• If the user asks for "daily/latest" without a date, default to today.
• If the user asks for a range, convert it to inclusive YYYY-MM-DD dates.

1. Discover candidate articles with scripts/article_discovery.py.
• Default sources live in references/sources.json.
• Use --today for daily mode.
• Use --since YYYY-MM-DD --until YYYY-MM-DD for a custom time window.
• Use --source all unless the user names a specific source.

1. Verify and read the strongest candidates.
• Open the source pages or article URLs with web browsing when available.
• Prioritize primary sources over reposts or commentary.
• If dates conflict between listing pages and article metadata, report the conflict and prefer explicit article metadata.

1. Analyze with separation between facts and judgment.
• Summarize only claims supported by the article.
• Identify the article's core thesis, novelty, evidence, assumptions, incentives, and missing context.
• Add your own take as a clearly labeled viewpoint, not as source fact.
• Compare against nearby articles when the user asks for trends or multiple sources.

1. Cite sources.
• Include article links for factual claims.
• Avoid long verbatim quotations; paraphrase unless a short quote is necessary.

1. Build the HTML report.
• Create an analysis JSON file in work/ with the shape described below.
• Run scripts/build_article_report.py --input work/analysis.json --output .
• Open or inspect the generated HTML enough to ensure the article navigation and sections are present.

Script Usage

Run from the skill directory or pass the config path explicitly:

python scripts/article_discovery.py --today
python scripts/article_discovery.py --since 2026-06-01 --until 2026-06-09
python scripts/article_discovery.py --source langchain --days 7 --limit 20
python scripts/article_discovery.py --source all --since 2026-06-01 --output work/articles.json
python scripts/build_article_report.py --input work/analysis.json --output outputs/article-intelligence-2026-06-09-0930.html

Analysis JSON Shape

Create this JSON after article discovery and reading. Keep values concise but substantive.

{
  "title": "AI Agent Article Intelligence",
  "generated_at": "2026-06-09 09:30",
  "date_range": "2026-06-09 to 2026-06-09",
  "summary": "No articles were published today; nearest relevant articles are included.",
  "trend_judgment": "Cross-article trend judgment.",
  "articles": [
    {
      "id": "stable-slug",
      "title": "Article title",
      "source": "LangChain Blog",
      "published_date": "2026-06-05",
      "url": "https://example.com/article",
      "relevance": "High",
      "one_line_thesis": "Short thesis.",
      "analysis": {
        "thesis": "Main thesis.",
        "key_claims": ["Claim 1", "Claim 2"],
        "evidence": ["Evidence or support."],
        "assumptions": ["Assumption."],
        "missing_context": ["Missing context."],
        "implications": ["Practical implication."]
      },
      "personal_take": "Clearly labeled informed opinion.",
      "original_notes": "Clean article notes, short excerpts, or paraphrased original-content map."
    }
  ]
}

The script returns JSON records:

{
  "source": "langchain",
  "title": "Example title",
  "url": "https://...",
  "published_date": "2026-06-05",
  "date_confidence": "metadata",
  "matched_from": "article"
}

If discovery returns no results, report no matches for the requested window and still generate the HTML report. Some blogs publish with local time, delayed index updates, or metadata that differs from visible dates; mention that only as a caveat, not as a reason to silently broaden the range.

Chat Summary Template

Keep the final chat response short because the HTML is the main artifact:

已生成 HTML 报告：[article-intelligence-YYYY-MM-DD-HHMM.html](...)

今天是否有新文章：...
本次纳入分析：...
我的核心判断：...

Quality Bar

• Prefer precision over volume; if there are many articles, rank the most important ones and say what was omitted.
• Do not treat marketing claims as proven outcomes.
• Watch for product-announcement bias, benchmark cherry-picking, and missing deployment constraints.
• In opinion sections, be concrete: name who benefits, what changes in practice, and what uncertainty remains.
• Use the user's language for the final answer unless they request otherwise.