Cited Dossier Verifier

SKILL: Cited Dossier Verifier
AI summaries hallucinate, which is fatal for due diligence. This grabs an AI-mode dossier with its citations, then re-queries each cited domain and dated press coverage — and optionally Reddit and X for independent first-hand corroboration — to confirm or quarantine every claim before it lands in your file. You control which corroboration channels run, the subject's jurisdiction, and how far back the press window reaches.

You are running this skill on API Direct via its MCP tools. Execute the steps below yourself by calling the named tools in order — values in <angle brackets> come from a previous step. Then deliver the result described at the end.

INPUTS:
  - subject (required): <subject — ASK THE USER>
      The person or company the dossier is about
  - focus (optional): (optional — e.g. regulatory actions and litigation)
      What to emphasize in the dossier
  - platforms (optional): (optional — e.g. reddit, twitter)
      Comma-separated list of platforms to run — options: web, news, reddit, twitter. Omit to run all of them; name specific platforms to limit the run.
  - country (optional): (optional — e.g. US)
      Two-letter country code to bias the AI dossier and news toward the subject's home jurisdiction (maps to the country param on google_ai_mode and search_news). Useful for jurisdiction-specific regulatory and legal claims.
  - news_window (optional): 1y (default — override if you like)
      How far back to pull dated press coverage for corroboration (maps to search_news time_published). One of 1h, 1d, 7d, 1y, anytime. Defaults to 1y.

PLATFORM SELECTION: some steps are gated with "Only run if {platforms} includes X". If the user supplied a `platforms` value, run only the steps whose platform is listed. If the user did NOT supply `platforms`, run ALL steps — every platform the skill supports.

STEPS:
  1. Tool `google_ai_mode` — google_ai_mode(prompt="Summarize controversies, lawsuits, regulatory actions and key facts about {subject}, focused on {focus}")
     Capture each claim from reply_parts and the supporting reference_links so every assertion has a candidate citation. Pass {country} to bias the answer toward the subject's home jurisdiction when set.
  2. Tool `search_web` — search_web(query="{subject}", include_ai_overview=true)
     Run a second independent AI overview plus organic results; compare its ai_overview.reference_links against the first pass to catch contradictions.
  3. Tool `search_web` — search_web(query="site:<citation_domain> {subject}")
     For each cited domain from steps 1 and 2, query the original source directly to confirm the claim is actually stated there and not fabricated.
  4. Tool `search_news` — search_news(query="{subject}", time_published=1y, limit=30)
     Corroborate or contradict the AI-stated claims against dated press coverage, flagging anything unsupported. 1y defaults to 1y; set {country} for jurisdiction-specific coverage.
  5. Tool `search_reddit` — search_reddit(query="{subject}", sort_by=relevance)
     Only run if {platforms} includes reddit: independently corroborate or contradict each AI claim against community discussion, surfacing first-hand accounts and contradictions the press missed before promoting any claim.
  6. Tool `search_twitter` — search_twitter(query="{subject}", sort_by=relevance)
     Only run if {platforms} includes twitter: cross-check claims against real-time public statements from the subject, accusers, and journalists; quarantine any claim no independent voice supports.

DELIVER: A fully-cited dossier where every claim is traced to and confirmed against its primary source plus any independent social corroboration you enabled, with unverifiable assertions clearly quarantined.

Note: each underlying tool call is billed at its normal endpoint price; get_sentiment adds a small per-page surcharge. Page through results as needed but stop once you have enough to deliver the outcome.