"An answer without a trail is a guess in formal clothes."
「沒有來源軌跡的答覆，是穿著正裝的猜測。」
An Agent Skill that forces Claude to trace every factual claim back to a primary source — and to disclose what's verified, what's uncertain, and what's a guess.
一個 Agent Skill，強制 Claude 把每一個事實宣稱上溯到原始來源，並揭露什麼已驗證、什麼不確定、什麼是猜測。
Unlike asking AI a factual question and getting a confident-sounding answer, this skill runs a structured verification workflow that classifies sources, fetches primary documents, labels confidence levels, and adversarially checks the conclusion before answering.
不同於問 AI 一個事實問題然後拿到一個聽起來很有自信的答案，這個 Skill 會執行結構化的查證流程：來源分級、抓取原始文件、標註信心等級、敵對審查結論，然後才回答。
When triggered on a factual question, the skill walks through a 6-phase verification workflow:
當在事實性問題上被觸發時，Skill 會走完 6 個查證階段：
| Phase | Action / 動作 |
|---|---|
| 0 | Trigger check — is this factual? / 觸發判斷：這是事實性問題嗎 |
| 0.3 | Key Assumptions Check — audit the question's hidden premises before searching / 前提核查（CIA Tradecraft Primer KAC 四步法） |
| 0.5 | Pre-register search strategy / 預先登記搜尋策略（防止事後挑詞） |
| 1 | Scope reconnaissance / 範圍偵察（Gemini 偵察兵 or broad WebSearch） |
| 2 | T1/T2 verification via WebFetch — with T1 Domain Gate / 抓取原始來源驗證（含 T1 域名強制規則） |
| 3 | Disclosure output / 揭露式呈現（信心標籤 + 來源紀錄 + PII discipline） |
| 4 | Peer review / 同儕審查 |
| 5 | Devil's advocacy + required ACH table / 敵對審查 + 必出競爭假說表 |
| 6 | Belief Source-Trail — explain why the false belief was plausible / 信念溯源 — 解釋錯誤指控為什麼自然流傳 |
Phases 1, 4, 5, 6 are conditional on stakes/context. Phases 0, 0.3, 0.5, 2, 3 always run.
Phase 1、4、5、6 視風險等級與情境條件執行。Phase 0、0.3、0.5、2、3 永遠執行。
Every source is classified before use. Factual claims must trace to T1 or T2.
每個來源使用前先分級。事實宣稱必須能上溯到 T1 或 T2。
| Tier | Type | Examples |
|---|---|---|
| T1 | Primary / Original 原始來源 | Official documents, raw data, party-of-record statements |
| T2 | Authoritative Secondary 權威次級 | Reuters, AP, peer-reviewed journals |
| T3 | Reported with citation 帶引用報導 | News articles citing primary sources |
| T4 | Aggregated summary 聚合摘要 | Wikipedia, summary sites |
| T5 | LLM-synthesized 大模型合成 | Claude / Gemini / GPT outputs — not evidence |
LLM agreement is not verification. Two models trained on similar corpora reaching the same conclusion adds little epistemic value.
LLM 之間的同意不是驗證。 兩個在類似語料上訓練的模型得出同一結論，認知價值很低。
Knowledge cutoff is a tool-call signal, not a give-up clause.
知識截止不是放棄的出口，是觸發工具的訊號。
When the model detects a time point past its training cutoff, the correct action is to call WebSearch / WebFetch immediately — not to disclaim and recommend the user check elsewhere. This is the single rule that does the most work.
當模型偵測到訓練截止之後的時間點，正確動作是立刻呼叫 WebSearch / WebFetch——而不是聲明限制然後叫使用者自己去查。這是整個 Skill 中影響力最大的單一規則。
This framework is a practical synthesis of established source-evaluation methodologies:
本框架整合三個學科的成熟方法論：
-
Source Criticism (Quellenkritik) — Leopold von Ranke, 19th-century historiography — Foundational method requiring every historical claim to trace back to primary documents. The T1-T5 tier system is its digital-age extension.
-
來源批判（Quellenkritik）— Ranke 19 世紀史學方法：要求每個歷史宣稱都必須能上溯到原始檔案。本 Skill 的 T1-T5 分級是其數位時代延伸。
-
Admiralty Code / NATO Source Reliability Rating — Military intelligence's two-axis system (A-F source reliability, 1-6 information credibility). The direct ancestor of T1-T5 classification.
-
Admiralty Code / NATO 情報界源可靠度評級：A-F 來源可靠度與 1-6 資訊可信度雙軸系統，是 T1-T5 的直系祖先。
-
Estimative Language (US Intelligence Community Directive 203) — Calibrated confidence labels (Verified / Likely / Possible / Speculation) used in intelligence reports to prevent confidence inflation.
-
Estimative Language（美國情報總監辦公室 ICD 203）：已驗證 / 高信心 / 中信心 / 低信心 / 不採信的標籤系統，源自美國情報界用以避免信心高估的標準。
git clone https://github.com/LionHuang/source-trail.git ~/.claude/skills/source-trailRestart Claude Code (or start a new session) for the skill to load.
重啟 Claude Code（或開新 session）讓 Skill 載入。
-
Download this repository as a ZIP (Code → Download ZIP)
-
Go to Settings → Features → Custom Skills
-
Upload the ZIP file
-
The skill activates automatically on factual questions — works on mobile too
-
下載本 repo 為 ZIP（Code → Download ZIP）
-
進入 Settings → Features → Custom Skills
-
上傳 ZIP
-
Skill 會在事實性問題出現時自動啟動——手機 App 也能用
Copy the SKILL.md and examples/ folder to your platform's skill directory. The Agent Skills format is compatible with 20+ coding agents.
複製 SKILL.md 和 examples/ 資料夾到對應平台的 skill 目錄。Agent Skills 格式相容於 20+ 個 coding agents。
For Phase 1 (scope reconnaissance) and Phase 4 (peer review) enhancements:
npm install -g @google/gemini-cli
gemini authIf Gemini is not installed, the skill gracefully degrades:
- Phase 1 → uses broad WebSearch instead of Gemini scout
- Phase 4 → uses Claude self-roleplay instead of Gemini peer review
未安裝 Gemini 時，Skill 會優雅降級：Phase 1 改用 broad WebSearch，Phase 4 改用 Claude 自我角色扮演批評者。
The skill auto-triggers on factual questions matching its description. You can also invoke manually with /source-trail.
Skill 會在符合描述的事實問題上自動觸發，也可手動 /source-trail 呼叫。
Examples / 範例：
2026 年台積電最新一季營收是多少？
What's OpenAI's latest model release this month?
Faker 現在還是韓國 LCK 最強的中路選手嗎？
Which Python ORM has the highest GitHub stars right now?
🔍 /source-trail activating
Stakes: Low
Expected: 2-3 searches
**2026 Q1：台積電淨收入 NT$1.13 兆，YoY +35%。**
毛利率 66.2%（歷史新高），每股淨利 $18.16 USD。
---
## 查證細節
**來源**: ir.tsmc.com/investor/quarterly-results (T1, 2026-04-17)
**反證條件**: 台積電若重述財報，此答案需更新。
**搜尋紀錄**: "TSMC 2026 Q1 quarterly results" / 2026-05-08
Verified facts are presented inline without labels (default = verified). Only non-verified exceptions carry ⚠ warnings. Phase numbers are never shown to the user.
已驗證事實直接內文呈現，不加標籤（預設 = 已驗證）。只有非驗證的例外會帶 ⚠ 警示。階段編號永遠不暴露給使用者。
The examples/ directory documents the failure modes that motivated each rule:
examples/ 目錄記錄了催生每條規則的失敗案例：
case-tsmc.md— Before/after the Iron Law: same skill structure produced "please check yourself" before, vs. actually fetching T1 data after. The clearest demonstration of the skill's net value.case-baseball.md— Phase 1 Gemini scouting succeeds where solo WebSearch fails (2026 Taiwanese MLB players, otherwise missed entirely).case-faker.md— Phase 4 peer review's double-edged nature: Gemini's methodological critique is valid AND fabricates its own specifics.case-phase-0-3-genesis.md— Meta-case: how Source-Trail verified its own academic foundations (CIA KAC, Popper, Maxwell, Kahneman) before adding Phase 0.3 to the skill.case-citation-laundering.md— When news outlets are mistakenly labelled T1: a 2026-05-11 failure during a legislative-bill comparison that introduced the T1 Domain Gate rule into Phase 2.case-political-substitution.md— When a viral question smuggles a false Authority premise: a 2026-05-11 case (Hualien "300 億" controversy) that hardened Phase 0.3 from declaration-based to execution-based.case-privacy-leak.md— When methodological rigor is not the same as ethical adequacy: a 2026-05-11 PII failure in the Skill's own examples that introduced the Privacy / PII Discipline rule into Phase 3.case-skipped-ach.md— When Phase 5 is run but ACH is silently skipped: a 2026-05-11 failure that hardened Phase 5's Analysis of Competing Hypotheses from a listed step to a required written table with at least H1 + one non-trivial alternative.case-belief-source-trail-origin.md— Why Phase 6 was added: factually correct verifications without an empathy layer get rejected. The complement to fact source-tracing — explaining why a false belief became plausible enough to spread, without dismissing the believer. "Emotions and anxieties are real; they must be empathized with before someone can read the source trail."
I use Claude as a research assistant. One day I asked it which Taiwanese MLB players had played the most games in 2026 — it confidently synthesized a wrong answer from outdated search snippets. Then I asked about TSMC's latest quarterly revenue — it bailed out with "my training data is from 2025, please check the IR website yourself." Then I asked if Faker was still Korea's strongest mid-laner — it answered half from training data and pushed the rest onto me.
我把 Claude 當研究助手用。有一天我問它 2026 年出賽最多的台籍 MLB 球員——它從過時的搜尋摘要合成了一個錯誤答案，還煞有介事地給我。我又問台積電最新一季營收——它說「訓練資料截止在 2025，建議你自己上 IR 網站查」就放棄了。我再問 Faker 現在還是韓國最強的中路選手嗎——它一半用訓練資料回答，一半把工作丟回給我。
These three failures are different surface symptoms of the same root cause: LLMs do not have epistemic discipline by default. They can call tools, but they don't always know when they should — they default to "answer from what I know" until you force them otherwise.
這三個失敗是同一個根本原因的不同表面症狀：LLM 預設沒有認識上的紀律。 它們會用工具，但不一定知道什麼時候該用——預設行為是「用知道的回答」，除非強制要求。
So I synthesized methodologies from journalism (multi-source verification), research ethics (pre-registration, peer review), intelligence tradecraft (estimative language, devil's advocacy), and evidentiary standards (chain of custody) into one skill that turns this discipline into Claude's default behavior on factual questions.
於是我把新聞工作（多重來源驗證）、研究倫理（預先登記、同儕審查）、情報分析（信心標籤、敵對審查）、法庭證據（證據鏈）這四個領域的方法論整合成一個 Skill，把這種紀律變成 Claude 在事實性問題上的預設行為。
This is the skill I wished Claude had by default.
這就是我希望 Claude 預設就有的 Skill。
Contributions welcome. The skill evolves based on observed failure modes:
歡迎貢獻。Skill 隨觀察到的失敗模式演進：
- Document a failure case in
examples/(skill triggered but produced wrong/unsafe output) - Propose a rule update in
SKILL.mdthat would have prevented it - Open a PR with both
When proposing rules, prefer relative descriptions over absolute values (e.g., "knowledge cutoff" not "August 2025"). Rules with hardcoded dates expire silently.
提議規則時，優先使用相對描述而非絕對數值（例如用「知識截止」而非「2025 年 8 月」）。寫死日期的規則會無聲過期。
MIT — see LICENSE.
Built by Lion Huang — where information rigor meets AI-augmented research.
由 Lion Huang 出品——資訊嚴謹度與 AI 輔助研究的交會點。