Agent Fact Check Verify

嚴謹多來源資訊查核與可信度判定技能。用於「查證/核實/核實這個/是真的嗎/是否正確」類請求，整合政府、官方、主流媒體、事實查核站、X(Twitter)、Reddit 等來源，採用內部 100 分制規則化評分（不對使用者公開分數），對外輸出中立且整合式結論。

Security Scan

The skill's code, instructions, and requirements are consistent with a multi‑source fact‑checking tool and do not request unrelated credentials or install arbitrary software.

Purpose & Capability

The name/description (multi‑source fact checking) aligns with the included files and runtime instructions. The Python script implements claim extraction, deterministic scoring, and user output composition described in SKILL.md. No unrelated credentials, binaries, or config paths are required.

Instruction Scope

Runtime instructions are scoped to: extract claims, run searches (agent tooling), optionally call twitter/reddit CLIs, score evidence, and compose an integrated reply. The skill does not instruct reading arbitrary user files or secrets. Caveat: SKILL.md documents optional use of X/Twitter and Reddit CLIs that rely on cookie/session files and suggests using the user's preferred search tool; if the agent is given cookie files or broad browser access, those external assets could be accessed — this is an operational risk but is explicit and optional in the docs.

Install Mechanism

There is no install spec (instruction-only with a bundled Python script). Nothing is downloaded from external URLs during install. The runtime script is local and self-contained.

Credentials

The skill requires no declared environment variables or primary credential. SKILL.md references an optional TAVILY_API_KEY and optional cookie/session files for social‑platform CLIs; these are optional and documented. Users should not provide API keys or cookie files unless they understand the implications (those artifacts grant access to accounts/sessions).

Persistence & Privilege

The skill does not request always:true, does not modify other skills or system settings, and has normal autonomous invocation settings. It writes/reads workspace files (claims, evidence, outputs) which is expected for this workflow.

Assessment

This skill appears internally consistent for multi‑source fact checking and is implemented by a local Python script that only processes evidence and composes replies. It does not require credentials by default. However, optional features (Tavily API key, twitter-cli/rdt-cli) use API keys or browser cookie/session files — only provide those if you trust the environment and understand that cookies/API keys grant access to the associated accounts. Also confirm your agent's search/browser tooling will perform network queries in ways you find acceptable, and avoid giving the skill unnecessary access to private or system files.

核心原則

• 對外回覆不展示評分機制與分數。
• 對內可拆 claim，但對外禁止逐條 claim 展示。
• 對外只提供整合後結論，語氣中立，不帶立場。
• 如為錯誤資訊，直接給出正確情形。
• 如為預測資訊，不做真偽裁定，僅整理目前可查情形。
• 最後永遠附上限制聲明：
  • ⚠️ 本查核基於公開可得資訊，無法涵蓋未公開或付費牆後的內容。

對外輸出格式（強制）

輸出時固定使用以下四段，順序不可更動：

1. 是否正確（簡答）：僅可用「正確 / 錯誤 / 部分正確 / 證據不足」其一，並附一句簡答。
2. 此事的真實情形：整合後敘述，不列逐條 claim。
3. 結論：給出可執行的最終判斷（必要時含不確定性提醒）。
4. 相關連結（最多三個）：最多 3 條，優先官方與高可信來源。

執行流程（Agent）

1. scripts/factcheck_engine.py extract：拆解可查證 claim。
2. 依 claim 類型分流：opinion/satire/prediction/verifiable。
3. 搜尋分三輪：官方→主流→反證（每輪最多 2 次，總 6 次）；另外可增加 X(Twitter) 多次查核（建議 3 次）作為社群佐證，且不改動官方/主流/反證三輪次數。
4. scripts/factcheck_engine.py score：以規則表計算內部分數。
5. scripts/factcheck_engine.py compose：輸出對使用者的整合回覆（不含分數）。

搜尋工具策略

• 預設優先使用使用者常用搜尋。
• 可選 Tavily（若環境有 TAVILY_API_KEY）。
• 可選 X/Reddit CLI：
  • X: twitter-cli（https://github.com/jackwener/twitter-cli）
  • Reddit: rdt-cli（https://github.com/jackwener/rdt-cli）
• CLI 未安裝時自動降級為一般網路搜尋，不中斷流程。

來源分級與評分

詳見：

• references/scoring-rubric.md
• references/source-policy.md

版本

• 1.0.4