med-news-reporter
把逐字稿和素材寫成新聞稿，不准捏造任何沒來源的事實
中文翻譯
---
name: "med-news-reporter"
description: "使用者想把原始素材（逐字稿、訪談、活動筆記、數據、直接引述）轉成可發表的新聞稿（即時新聞、調查報導、特稿或評論）時使用。啟動完整的新聞編輯室工作流程：類型選擇、素材盤點、事實查核、平衡、媒體倫理紅線，以及媒體識讀自我檢查。也會在這些說法出現時觸發：'write up this transcript'、'turn into a news article'、'organize into a feature'、'polish into a report'、'draft an op-ed'、'幫我寫成新聞稿'、'潤成一篇報導'、'整理成專訪'、'寫一篇關於 X 的評論'、'把逐字稿做成 feature'，就算使用者沒有說「新聞」這兩個字也一樣。不要用在：新聞稿／公關稿（用 pr-press-release）或行銷文案（用 mkt-*）。"
metadata:
category: "WP-50 大眾傳播"
tags: ["news", "journalism", "reporting", "media-ethics", "media-literacy"]
---
# 新聞記者：專業新聞寫作
## 概述
濃縮自 26 所新聞學院（政大、台大、Columbia、Missouri、Medill、UC Berkeley、Sciences Po、中文大學、港大 JMSC 等）的寫作流程。把原始素材（逐字稿、數據、活動筆記、直接引述）轉成可發表的稿件，涵蓋四種經典類型：即時新聞、調查報導、特稿、評論。
```
IRON LAW: No Unsourced Facts
Every concrete fact in the finished piece — names, titles, numbers, dates,
places, quotes, causal claims — must trace to material the user provided.
Anything missing is marked [待查證: specific description] in the draft,
NOT silently filled in with plausible-sounding invention. This holds even
when the gap is small ("probably around 30%", "most likely Tuesday"):
either the source exists or the placeholder stays. A fabricated plausible
detail is a defamation / retraction / trust-collapse vector — treat it as
radioactive. When in doubt, ask the user before writing, not after.
```
為什麼這件事不直覺：LLM 預設會「補上」資訊讓文字順暢（一個合理的職稱、一個整數、一句修順的引述）。在新聞業，這是已發表的不實內容最常見的來源。鐵律就是用來壓住這個預設行為的。
**合理化對照表：以下這些說法都「不能」推翻鐵律：**
| Claude 可能會想… | 為什麼這樣還是違規 |
|---|---|
| 「使用者說『大約 400』，我進位成 500，文字比較乾淨」 | 任何憑空加上的精確度都是捏造。用提供的數字，不然就用 [待查證]。 |
| 「這個市場規模大家都知道，不需要來源」 | 就算是公開知識，也要能追溯到使用者提供的素材，不然就要標記。 |
| 「我只是重建引述的意思，不是確切的字句」 | 用轉述加註出處；絕對不要重建成直接引述。 |
| 「這個缺口很小，不重要」 | 重不重要是編輯要判斷的，不是寫稿的人。佔位標記要留著。 |
| 「補一個合理的分析師估計，讓報導更完整」 | 捏造專家意見是 LLM 新聞寫作最常見的失敗。不行。 |
---
## 何時使用
**觸發條件：**
- 使用者提供原始素材（逐字稿、訪談、活動筆記、數據、新聞稿、紀錄），並要求把它變成類似新聞的稿件。
- 使用者要求「news article」、「report」、「feature」、「op-ed」、「column」、「commentary」、「investigation」、「profile」、「long-form story」、「稿件」、「深度報導」、「專稿」、「評論」。
- 使用者換句話說：「幫我寫成新聞稿」、「潤成一篇報導」、「整理成專訪」、「寫一篇關於 X 的評論」、「把逐字稿做成 feature」。
- 使用者要求「把公關新聞稿改寫成新聞」（具體來說：拿公關素材，產出新聞語氣的報導）。
**輸入訊號：**
- 素材裡有具名的人、組織、時間、地點、數字或直接引述，也就是可能出現在真實新聞報導裡的內容。
- 有明確或隱含的發表意圖（「給編輯室用」、「我們要發表這篇」、「要投稿到 X 媒體」）。
- 要求裡提到中立、平衡、出處標註或事實查核。
**何時不要用：**
- 當事公司以自家口吻發的新聞稿或公告 → 用 `pr-press-release`。
- 行銷文案、廣告創意、landing page 文案 → 用 `mkt-*`。
- 社群貼文草稿（短篇、宣傳性質）→ 用 `pr-social-copywriting`。
- 沒有發表意圖的內部會議紀錄／摘要 → 用 `ops-meeting-minutes`。
---
## 方法
### Step 1：判斷稿件類型
從四種類型挑一個，然後讀對應的參考檔：
| 類型 | 訊號 | 讀這個 |
|------|---------|------|
| **即時新聞／硬新聞** | 事件、記者會、公告，5W1H 齊全 | `references/type_breaking_news.md` |
| **調查報導／深度報導** | 多方來源、隱藏的事實、系統性問題、文件交叉比對 | `references/type_investigative.md` |
| **特稿／長篇敘事** | 人物側寫、場景、敘事弧線、以主題驅動 | `references/type_feature.md` |
| **評論／專欄／op-ed** | 立場、詮釋、論證 | `references/type_opinion.md` |
如果不明確，**問使用者**，不要猜。如果素材橫跨多種類型，就依使用者指定的類型。
### Step 2：素材盤點與找出缺口
起草之前：
1. **列出手上有的事實**：素材裡的人物、時間、地點、事件、數字、引述。
2. **標記來源強度**：
- 一手（逐字稿、原始文件、現場筆記）→ 可直接使用。
- 二手（其他媒體轉述、轉傳訊息）→ 必須交叉查證。
- 傳言／沒有來源 → 不能用，或明確標示為「據稱」／「未經證實」。
3. **標記缺口**：5W1H 缺了哪個？有沒有對立方？數字有沒有來源？引述有沒有前後文？**每個缺口都要在稿子裡以 `[待查證: specific description]` 浮現，或是一開始就問。**
如果來源強度或查證方式不清楚，參考 `references/fact_checking.md`。
### Step 3：套用類型範本
> **鐵律檢查：** 只能用 Step 2 素材盤點中確認過的名字、數字和引述。任何缺少的事實 → `[待查證: description]`，不要自己補。
依 Step 1 載入的參考範本寫作。跨類型通用的原則：
- **導言（Lead）**：30–50 字；即時新聞用倒金字塔，特稿可以用場景／人物／提問式導言。
- **出處標註**：直接引述 `「…」王小明說。` ／ 間接引述 `王小明表示…`。**不要改變引述的意思**；整理標點可以。
- **匿名來源**：只有在 (a) 消息來源面臨真實風險，而且 (b) 沒有其他選擇時才用。要說明理由（「為避免遭報復而要求匿名」）。
- **數字**：一律要註明來源。比例要搭配絕對數字（「裁員 30%，約 1,200 名員工」）。避免誤導性的比較。
- **平衡**：報導指控或爭議時，**要給被指控的一方回應的機會**。如果對方拒絕或聯絡不上，要寫出來（「多次聯繫 X 公司；截至發稿時未獲回應」）。
- **揭露**：業配內容、聯盟行銷利益、利益衝突，都要在文末揭露。
寫作風格的預設：
- **具體勝過抽象**：寫「月薪 3 萬 2 千元」，不要寫「薪資不高」。
- **動詞勝過形容詞**：「抨擊」／「質疑」／「譴責」比「嚴重地反對」好。
- **主動勝過被動**：「警方逮捕嫌犯」，被動只在要強調受詞時才用。
- **短句**：平均 ≤ 40 字；連續超過 3 個複合句就要警覺。
- **禁止**：沒有來源的讀心（「他心中十分憤怒」→ 改寫成行為：「他拍桌大聲表示…」）；過度渲染的形容詞（「令人震驚」）；立場外洩（「正義的警方終於逮到嫌犯」）；傳聞（「聽說」／「有人說」）。
### Step 4：媒體倫理檢查（必做）
**輸出之前，你必須把 `references/media_ethics.md` 裡下面每一項都做完。** 不要因為「這篇看起來很乾淨」就跳過，因為這個清單就是要推翻這種判斷。
1. 有誹謗風險嗎？（對具名的真實人物做出未經查證的負面宣稱）
2. 有侵犯隱私嗎？（未經同意揭露私人資訊）
3. 消息來源有受到保護嗎？（匿名來源能不能從細節被重新辨識出來？）
4. 是特殊類別的題材嗎？（未成年人、性侵被害人、自殺，台灣法律有特定限制）
5. 有沒有未揭露的利益衝突？
6. 圖片／素材的授權？
**只要有任何一項不確定，就要在輸出裡明確警告使用者。** 不要默默放行。
### Step 5：媒體識讀自我檢查（必做）
**輸出之前，你必須把 `references/media_literacy.md` 裡下面每一項都做完。** 一篇「感覺很平衡」的稿子是最常見的失敗模式，這個清單抓的就是直覺漏掉的東西。
1. 導言有沒有誇大（標題黨）？標題和內文對得上嗎？
2. 事實和意見有沒有混在一起？事實性的宣稱用陳述語氣；意見用轉述語氣（「批評者認為」、「專家表示」）。
3. 這篇稿子是在訴諸情緒而不是證據嗎？
4. 數據的呈現有沒有可能誤導（基期、挑選過的時間區間、把相關當因果）？
5. 有沒有標出來源層級（官方／當事人／第三方／匿名）？
6. 如果有 AI 協助產生或整理任何內容，有沒有揭露？
### Step 6：輸出成稿
依下面的輸出格式輸出。如果使用者明確要純稿件、不要 meta footer，就省略 footer，但 Step 4–5 還是要在內部做完。
遇到以下情況，**停下來問使用者**：
1. 素材缺少 5W1H 當中的 2 項以上。
2. 素材涉及未成年人、性侵害、自殺，或資訊不完整的醫療題材。
3. 對另一方的指控只有單一來源。
4. 沒有指定類型，而素材橫跨多種類型。
5. 素材有矛盾（同一件事有兩個版本）。
6. 要求的篇幅跟素材量不符（例如用 200 字的簡報要寫 3000 字的調查報導）。
所有缺口一次問完，不要來回問。
---
## 輸出格式
```markdown
# [Headline: ≤ 20 chars, concrete people/events]
**副標**: (optional, 15–25 chars)
[Lead paragraph]
[Body paragraphs…]
---
**稿件類型**: 即時新聞 / 深度調查 / 特稿 / 評論
**字數**: approx. XXX
**消息來源層級**: 一手訪談 N 則 / 二手引用 N 則 / 匿名 N 則
**待查證事項**:
- [ ] [specific item]
**倫理／識讀檢核摘要**:
- 平衡原則: ✅ / ⚠️ (reason)
- 匿名來源揭露: ✅ / N/A
- 利益揭露: ✅ / N/A
- 情緒化字眼: ✅ / ⚠️ (list)
- 數據來源: ✅ / ⚠️ (list)
```
---
## 範例
### 好的範例
**情境：** 使用者提供一份 1,200 字的記者會逐字稿，內容是一家台北餐飲 SaaS 新創宣布 A 輪募資。素材包含：執行長的引述、領投方合夥人的引述、募資金額（NT$120M）、目前客戶數（400+）、目標市場（日本）、具名的公司。
**分析：**
1. Step 1：歸類為即時新聞（事件驅動、5W1H 齊全）→ 載入 `references/type_breaking_news.md`。
2. Step 2：素材盤點：5W1H 完整；所有數字都能追溯到逐字稿；兩則直接引述都有正確標註出處；找到的缺口：沒有提供日本市場規模的數字。在稿子裡標為 `[待查證: 日本餐飲 SaaS 滲透率來源]`。
3. Step 3：倒金字塔導言：WHO（新創）+ WHAT（NT$120M A 輪）+ WHEN（今天）+ WHY（拓展日本）。後面的段落補上數據、投資人引述、產品細節。執行長的引述逐字使用。
4. Step 4：倫理檢查：沒有未成年人／被害人／私人資訊的風險；沒有未揭露的利益衝突；有一位投資人是大型創投（公開資訊）。
5. Step 5：識讀檢查：沒有標題黨；事實和投資人對未來的陳述分得清楚（「合夥人 X 表示團隊『有機會達到…』」）；數字都搭配來源。
6. 輸出包含 meta footer，標出那一個 `[待查證]` 項目和平衡評級。
結果：一篇可發表的硬新聞稿，新聞編輯只需要查證那一個標出來的數據點。沒有捏造的事實、沒有立場外洩、沒有倫理隱患。
### 壞的範例
**情境：** 輸入素材同上。寫稿的人為了讓稿子「更順」：(a) 把 400 個客戶進位成「超過 500」讓文字比較乾淨，(b) 把執行長的引述改寫成更有力的版本，(c) 加了一句「市場分析師預期日本擴張第一年可帶來 500 萬美元 ARR」，卻沒有提供任何分析師來源，(d) 因為「現在看起來很乾淨」就把 `[待查證]` footer 拿掉。
**哪裡出了問題：**
- (a) 捏造了更大的客戶數字，違反鐵律。就算是「進位」也是造假。
- (b) 修改引述會改變出處的歸屬。就算意思沒變，那也不是那個人說的話，有誹謗／準確性的風險。
- (c) 憑空捏造「分析師預期」，是典型的 LLM 幻覺模式。編輯會馬上把它抽掉，整篇稿子就死了。
- (d) 把稽核 footer 拿掉看起來比較漂亮，卻藏住了真正的信任問題：沒有 `[待查證]` 標記，下游的編輯根本沒有訊號知道 (c) 從來沒被查證過。
總結：讀起來很流暢，但這四個修改每一個都是會被開除等級的新聞錯誤。鐵律的全部意義就是擋下這些預設行為。
---
## 注意事項
- **「聽起來像真的引述」就是捏造的開端**：如果某句引述不是逐字稿裡的原話，就不要根據前後文「重建」它。用轉述加註出處，或標 `[引言待查證]`。重建的引述是新聞系學生犯錯最大宗的來源。
- **台灣的法律風險跟美國不同**：刑法 310（誹謗罪）即使陳述屬實，若缺乏公共利益也適用；偵查不公開限制報導進行中偵查的細節，就算記者知道也一樣；性侵害犯罪防治法禁止揭露可能辨識出性侵被害人的資訊。凡是碰到這些的內容，發表前先看 `references/media_ethics.md`。
- **WHO 自殺報導守則是強制的，不是選擇性的**：不要描述方式、地點，也不要刊登遺書。一定要附上求助專線的文字。很多編輯室的危機，都是因為「覺得有新聞價值」而跳過這條。
- **平衡 ≠ 虛假對等**：給對方回應的機會是必須的；在證據一面倒的情況下硬造一個「兩造各有道理」的框架，是在誤導。如果被指控的一方拒絕回應，就寫出來；不要用臆測的辯護來湊數。
- **匿名來源 ≠ 沒有出處的來源**：每個匿名來源都還是要在稿子裡看得到匿名的理由。「一位消息人士表示」卻沒有任何脈絡，是警訊，不是專業做法。
- **標題和內文不符是最快毀掉信任的方式**：如果標題承諾的比內文交代的多，就改標題，絕對不要把內文灌水。自我檢查時，最後再重讀標題，對照實際的內文。
---
## 參考資料
| 檔案 | 用途 | 何時讀 |
|------|---------|--------------|
| `references/type_breaking_news.md` | 即時新聞範本與範例 | Step 1：即時新聞 |
| `references/type_investigative.md` | 調查報導範本 | Step 1：調查報導 |
| `references/type_feature.md` | 特稿／敘事範本 | Step 1：特稿 |
| `references/type_opinion.md` | 評論／專欄範本 | Step 1：評論 |
| `references/media_literacy.md` | 自我檢查清單 | Step 5（一律要讀） |
| `references/media_ethics.md` | 倫理與台灣法律紅線 | Step 4（一律要讀） |
| `references/fact_checking.md` | 查證方法、平衡 | 來源層級不清楚時 |
相關 skill：`pr-press-release`（公司口吻的公告）、`pr-crisis-communication`（危機期間的訊息）、`hum-source-criticism`（更深入的來源審查框架）、`hum-ethics`（道德框架推理）。
---
## 限制
- **不會對照外部來源查證事實宣稱。** 這個 skill 檢查的是內部一致性並標出缺口；它不會對資料庫、法院紀錄或其他外部系統做即時查證。使用者必須自己做，或提供已查證的素材。
- **以台灣為中心的法律框架。** 媒體倫理與法律紅線引用的是台灣法律（刑法 310、個資法、偵查不公開、性侵害犯罪防治法、兒少法）。用在美國／歐盟／香港的司法管轄區時，一般原則仍適用，但具體法條需要替換。
- **超過約 5,000 字的長篇調查。** 對於書籍等級的專案、多篇系列報導或紀錄片處理，這裡的範本只涵蓋單篇稿件的工藝；跨系列的敘事架構需要這個 skill 以外的額外規劃。
- **不能取代法律審查。** 稿子裡有具名指控時，不管這個 skill 的倫理檢查有沒有過，不經律師審閱就發表還是有風險。
- **不會自己產生原創報導。** 如果提供的素材不夠，這個 skill 會標出缺口，但不會自己去找其他來源。英文原文
---
name: "med-news-reporter"
description: "Use when the user wants to turn raw material — transcripts, interviews, event notes, data, direct quotes — into a publishable news piece (breaking news, investigative report, feature, or op-ed). Activates the full newsroom workflow: type selection, material audit, fact-checking, balance, media-ethics red lines, and media-literacy self-check. Also triggers on phrases like 'write up this transcript', 'turn into a news article', 'organize into a feature', 'polish into a report', 'draft an op-ed', '幫我寫成新聞稿', '潤成一篇報導', '整理成專訪', '寫一篇關於 X 的評論', '把逐字稿做成 feature' — even when the user does not say the word 'news'. Do NOT use for press releases (use pr-press-release) or marketing copy (use mkt-*)."
metadata:
category: "WP-50 大眾傳播"
tags: ["news", "journalism", "reporting", "media-ethics", "media-literacy"]
---
# News Reporter — Professional News Writing
## Overview
Condensed writing workflow from 26 journalism schools (NCCU, NTU, Columbia, Missouri, Medill, UC Berkeley, Sciences Po, CUHK, HKU JMSC, ...). Turns raw material — transcripts, data, event notes, direct quotes — into a publishable piece across four canonical types: breaking news, investigative, feature, and opinion.
```
IRON LAW: No Unsourced Facts
Every concrete fact in the finished piece — names, titles, numbers, dates,
places, quotes, causal claims — must trace to material the user provided.
Anything missing is marked [待查證: specific description] in the draft,
NOT silently filled in with plausible-sounding invention. This holds even
when the gap is small ("probably around 30%", "most likely Tuesday"):
either the source exists or the placeholder stays. A fabricated plausible
detail is a defamation / retraction / trust-collapse vector — treat it as
radioactive. When in doubt, ask the user before writing, not after.
```
Why this is non-obvious: LLMs default to "filling in" to make prose flow (a reasonable title, a round number, a smoothed quote). In journalism this is the single most common route to published falsehood. The Iron Law suppresses that default.
**Rationalization Table — these justifications DO NOT override the Iron Law:**
| Claude might think... | Why it's still a violation |
|---|---|
| "The user said 'around 400', I'll round up to 500 for cleaner prose" | Any invented precision is fabrication. Use the supplied figure or [待查證]. |
| "This market size is widely known, I don't need a source" | Even public knowledge must trace to user-supplied material or be flagged. |
| "I'm just reconstructing the quote's meaning, not the exact words" | Paraphrase with attribution; never reconstruct as a direct quote. |
| "The gap is small and not important" | Importance is for the editor to judge, not the writer. Placeholder stays. |
| "Adding a plausible analyst estimate makes the story more complete" | Invented expert opinion is the most common LLM journalism failure. No. |
---
## When to Use
**Trigger conditions:**
- User supplies raw material (transcripts, interviews, event notes, data, press releases, logs) and asks for it to become a news-like piece.
- User asks for a "news article", "report", "feature", "op-ed", "column", "commentary", "investigation", "profile", "long-form story", "稿件", "深度報導", "專稿", "評論".
- User paraphrases: "幫我寫成新聞稿", "潤成一篇報導", "整理成專訪", "寫一篇關於 X 的評論", "把逐字稿做成 feature".
- User asks to "rewrite a PR release as news" (specifically: taking PR material and producing journalistic-toned coverage).
**Input signals:**
- Presence of named people, organizations, times, places, numbers, or direct quotations — i.e., material that could appear in a real news article.
- Explicit or implicit publication intent ("for the newsroom", "we're publishing this", "要投稿到 X 媒體").
- Requests mentioning neutrality, balance, attribution, or fact-checking.
**When NOT to use:**
- Press release or announcement from the subject company's own voice → use `pr-press-release`.
- Marketing copy, ad creative, landing-page copy → use `mkt-*`.
- Social-media post drafting (short form, promotional) → use `pr-social-copywriting`.
- Internal meeting minutes / summaries with no publication intent → use `ops-meeting-minutes`.
---
## Methodology
### Step 1: Classify the Story Type
Pick one of the four types, then read the matching reference:
| Type | Signals | Read |
|------|---------|------|
| **Breaking news / straight news** | Event, press conference, announcement, 5W1H available | `references/type_breaking_news.md` |
| **Investigative / deep report** | Multi-source, hidden facts, systemic issues, document cross-check | `references/type_investigative.md` |
| **Feature / long narrative** | Profile, scene, narrative arc, theme-driven | `references/type_feature.md` |
| **Opinion / column / op-ed** | Stance, interpretation, argument | `references/type_opinion.md` |
If ambiguous, **ask the user** — do not guess. If material spans multiple types, follow the user's specified type.
### Step 2: Material Audit & Gap Identification
Before drafting:
1. **List available facts**: people, times, places, events, numbers, quotes from the material.
2. **Tag source strength**:
- First-hand (transcript, original doc, on-site notes) → usable directly.
- Second-hand (other-media relay, forwarded message) → must cross-verify.
- Rumor / no source → unusable, or explicitly marked as "allegedly" / "unverified".
3. **Tag gaps**: which 5W1H is missing? Is there a counter-side? Do numbers have a source? Do quotes have context? **Every gap surfaces in the draft as `[待查證: specific description]` or is asked upfront.**
When source strength or verification method is unclear, consult `references/fact_checking.md`.
### Step 3: Apply the Type Template
> **Iron Law check:** only use names, numbers, and quotes confirmed in Step 2's material audit. Any missing fact → `[待查證: description]`, not fill-in.
Write per the reference template loaded in Step 1. Cross-type principles:
- **Lead**: 30–50 chars; breaking news uses inverted pyramid, features may use scene / character / question leads.
- **Attribution**: direct quote `「…」王小明說。` / indirect `王小明表示…`. **Do not alter quote meaning**; punctuation cleanup is OK.
- **Anonymous sources**: only when (a) the source faces real risk and (b) no alternative exists. State the reason ("requested anonymity to avoid retaliation").
- **Numbers**: always cite the source. Pair ratios with absolutes ("layoffs of 30%, about 1,200 employees"). Avoid misleading comparisons.
- **Balance**: when reporting an accusation or dispute, **give the accused a chance to respond**. If they refuse or are unreachable, state so ("reached X Company multiple times; no response by press time").
- **Disclosure**: sponsored content, affiliate interests, conflicts — disclose at the tail.
Writing-style defaults:
- **Concrete over abstract**: "月薪 3 萬 2 千元" not "薪資不高".
- **Verbs over adjectives**: "抨擊" / "質疑" / "譴責" beat "嚴重地反對".
- **Active over passive**: "警方逮捕嫌犯" — use passive only to emphasize the receiver.
- **Short sentences**: average ≤ 40 chars; >3 consecutive compound sentences is a warning.
- **Forbidden**: unsourced mind-reading ("他心中十分憤怒" → change to behavior: "他拍桌大聲表示…"); overcharged adjectives ("令人震驚"); stance leakage ("正義的警方終於逮到嫌犯"); hearsay ("聽說" / "有人說").
### Step 4: Media Ethics Check (required)
**YOU MUST complete every item below in `references/media_ethics.md` before producing the output.** Do not skip on the basis that "this piece looks clean" — that judgment is exactly what this checklist exists to override.
1. Defamation risk? (unverified negative claims about a named real person)
2. Privacy breach? (disclosing private info without consent)
3. Source protection? (can an anonymous source be re-identified from details?)
4. Special-category topic? (minors, sexual-assault victims, suicide — Taiwan law has specific restrictions)
5. Undisclosed conflict of interest?
6. Image / material licensing?
**If any item is uncertain, warn the user explicitly in the output.** Do not silently pass.
### Step 5: Media Literacy Self-Check (required)
**YOU MUST complete every item below in `references/media_literacy.md` before producing the output.** A piece that "feels balanced" is the most common failure mode — the checklist catches what intuition misses.
1. Does the lead overstate (clickbait)? Does the headline match the body?
2. Are facts and opinions mixed? Factual claims use declarative voice; opinions use reported voice ("critics argue", "experts say").
3. Is the piece appealing to emotion instead of evidence?
4. Could the data presentation mislead (base-rate, cherry-picked window, correlation-as-causation)?
5. Are source tiers marked (official / principal / third-party / anonymous)?
6. If AI helped generate or organize any content, is that disclosed?
### Step 6: Output the Finished Piece
Emit per the Output Format below. If the user explicitly wants a pure article with no meta-footer, omit the footer but still complete Steps 4–5 internally.
**Stop and ask the user** when:
1. 2+ of 5W1H are missing from material.
2. Material involves minors, sexual assault, suicide, or medical topics with incomplete info.
3. Only a single source exists for an accusation against another party.
4. Type is unspecified and material spans multiple types.
5. Material contains contradictions (two versions of the same fact).
6. Requested length mismatches material volume (e.g., 3000-word investigation from 200-word briefing).
Ask all gaps in one message, not back-and-forth.
---
## Output Format
```markdown
# [Headline: ≤ 20 chars, concrete people/events]
**副標**: (optional, 15–25 chars)
[Lead paragraph]
[Body paragraphs…]
---
**稿件類型**: 即時新聞 / 深度調查 / 特稿 / 評論
**字數**: approx. XXX
**消息來源層級**: 一手訪談 N 則 / 二手引用 N 則 / 匿名 N 則
**待查證事項**:
- [ ] [specific item]
**倫理／識讀檢核摘要**:
- 平衡原則: ✅ / ⚠️ (reason)
- 匿名來源揭露: ✅ / N/A
- 利益揭露: ✅ / N/A
- 情緒化字眼: ✅ / ⚠️ (list)
- 數據來源: ✅ / ⚠️ (list)
```
---
## Examples
### Good Example
**Scenario:** User provides a 1,200-word press-conference transcript from a Taipei restaurant-SaaS startup announcing Series A funding. Material contains: CEO quote, lead-investor partner quote, funding amount (NT$120M), current customer count (400+), target market (Japan), named companies.
**Analysis:**
1. Step 1: classified as breaking news (event-driven, 5W1H complete) → load `references/type_breaking_news.md`.
2. Step 2: material audit — 5W1H intact; all numbers trace to transcript; two direct quotes properly attributed; gap identified: no Japan-market sizing number supplied. Flagged as `[待查證: 日本餐飲 SaaS 滲透率來源]` in draft.
3. Step 3: inverted pyramid lead — WHO (startup) + WHAT (NT$120M Series A) + WHEN (today) + WHY (Japan expansion). Supporting paras add data, investor quote, product detail. CEO quote used verbatim.
4. Step 4: ethics check — no minors / victims / private-info risk; no undisclosed conflict; one investor name is a major VC (public info).
5. Step 5: literacy check — no clickbait; facts and investor's forward-looking statement clearly separated ("Partner X said the team 'could reach…'"); numbers paired with sources.
6. Output includes meta footer flagging the one `[待查證]` item and balance rating.
Result: publishable straight-news piece that a news editor would only need to verify the one flagged datapoint. No invented facts, no stance leakage, no ethics hazards.
### Bad Example
**Scenario:** Same input material as above. Writer produces a "smoother" piece by (a) rounding 400 customers to "over 500" for cleaner prose, (b) paraphrasing the CEO quote into a punchier version, (c) adding "market analysts expect Japan expansion to generate $5M ARR in year one" without any supplied analyst source, (d) omitting the `[待查證]` footer because "it looks clean now".
**What went wrong:**
- (a) fabricated a larger customer number — Iron Law violation. Even a "rounding up" is falsification.
- (b) quote modification changes attribution. Even if meaning is preserved, it's not what the person said — defamation / accuracy exposure.
- (c) invented an "analyst expectation" — a classic LLM hallucination pattern. An editor would pull this immediately and kill the piece.
- (d) removing the audit footer looks polished but hides the real trust problem: without `[待查證]` flags, a downstream editor has no signal that item (c) was never verified.
Net effect: reads fluent, but every one of those four edits is a sacking-level journalism error. The Iron Law's whole point is to block these defaults.
---
## Gotchas
- **"Sounds like a real quote" is how fabrication begins**: if a quote is not verbatim in the transcript, do not "reconstruct" it from context. Paraphrase with attribution, or mark `[引言待查證]`. Reconstructed quotes are the single biggest source of journalism-school failures.
- **Taiwan legal exposure differs from US**: 刑法 310 (誹謗罪) applies to true statements too if they lack public interest; 偵查不公開 restricts reporting details of ongoing investigations even when a reporter knows them; 性侵害犯罪防治法 forbids revealing information that could identify sexual-assault victims. See `references/media_ethics.md` before publishing anything touching these.
- **WHO suicide reporting guidelines are mandatory, not optional**: do not describe method, location, or publish a suicide note. Always include helpline text. Many newsroom crises come from skipping this because "it seemed newsworthy".
- **Balance ≠ false equivalence**: giving a chance to respond is required; inventing a "both sides" frame where the evidence is one-sided is misleading. If the accused declined to respond, state so; do not pad with speculative defenses.
- **Anonymous source ≠ unattributed source**: every anonymous source must still have a stated reason for anonymity visible in the piece. "A source said" with no context is a red flag, not professional practice.
- **Headline-body mismatch is the fastest trust-destroyer**: if the headline promises more than the body delivers, rewrite the headline — never inflate the body. In the self-check, re-read the headline last, against the actual body.
---
## References
| File | Purpose | When to read |
|------|---------|--------------|
| `references/type_breaking_news.md` | Breaking-news template & examples | Step 1: breaking news |
| `references/type_investigative.md` | Investigative template | Step 1: investigation |
| `references/type_feature.md` | Feature / narrative template | Step 1: feature |
| `references/type_opinion.md` | Opinion / column template | Step 1: opinion |
| `references/media_literacy.md` | Self-check list | Step 5 (always) |
| `references/media_ethics.md` | Ethics & Taiwan legal red lines | Step 4 (always) |
| `references/fact_checking.md` | Verification methods, balance | When source tiers are unclear |
Related skills: `pr-press-release` (for company-voice announcements), `pr-crisis-communication` (for crisis-period messaging), `hum-source-criticism` (for deeper source-vetting frameworks), `hum-ethics` (for moral-framework reasoning).
---
## Limitations
- **Does not verify factual claims against external sources.** This skill checks internal consistency and flags gaps; it does not perform live fact-checking against databases, court records, or other external systems. The user must do this or provide verified material.
- **Taiwan-centric legal framing.** Media ethics and legal red lines reference Taiwan law (刑法 310、個資法、偵查不公開、性侵害犯罪防治法、兒少法). For US / EU / HK jurisdictions, the general principles still apply but specific statutes need substitution.
- **Long-form investigations beyond ~5,000 words.** For book-length projects, multi-part series, or documentary treatments, the templates here cover single-article craft; narrative architecture across a series needs additional planning beyond this skill.
- **Not a substitute for legal review.** When the piece contains a named accusation, publishing without attorney review is still risky regardless of this skill's ethics pass.
- **Does not generate its own original reporting.** If the supplied material is insufficient, this skill flags the gap — it will not go find additional sources.