bible-fact-check（圣经事实核查）

Bible Fact Check

Systematically audit biblical reference content through 10 quality checks. Report issues with specific line numbers. Never auto-fix — only report.

Checks 1-5 apply to ALL content types (URLs, pasted text, files). Checks 6-10 apply to fun-facts reference files only — skip entirely for URLs and pasted text (do not print N/A lines for them).

Input Sources (pick one)

1. bible-buddy reference file — e.g., /review-bible-buddy fun-facts.md → reads from bible-buddy/references/
2. Any file path — e.g., /review-bible-buddy /path/to/my-notes.md → reads the specified file
3. URL — e.g., /review-bible-buddy https://example.com/article → Do NOT use WebFetch — many news sites block it with 403. Run uv run --directory {BIBLE_BUDDY} scripts/fetch_url.py "<URL>" to extract article text. The script auto-falls back to patchright (headless Chromium) when urllib gets 403.
4. Pasted text — user pastes content directly in the conversation → review the pasted text
5. No argument — use AskUserQuestion to ask the user whether they want to paste a URL or input text. Do NOT default to any file.

Path Resolution

Resolve {BIBLE_BUDDY} before anything else. Check in order:

1. Project-level: .claude/skills/bible-buddy/ (relative to repo root)
2. User-level: ~/.claude/skills/bible-buddy/

Use the first path where references/ directory exists. If neither exists, stop immediately and tell the user: 「bible-fact-check 需要 bible-buddy skill。請先安裝：

• 專案層級：npx skills add lancetw/skills/bible-buddy --project
• 使用者層級：npx skills add lancetw/skills/bible-buddy」

All {BIBLE_BUDDY} references below use the resolved path.

Prerequisites

1. Run dependency setup (one-time): uv sync --directory {BIBLE_BUDDY} && uv run --directory {BIBLE_BUDDY} patchright install chromium

How to Run

1. Determine the input source
2. Resolve {BIBLE_BUDDY} path (see Path Resolution above)
3. Read / receive the content

- For fun-facts.md: Do NOT use Read (exceeds 10K token limit). Use Grep: Grep("^- ", path="{BIBLE_BUDDY}/references/fun-facts.md", output_mode="content", head_limit=0) → returns all ~134 facts with line numbers in one call

1. Load reference files as checking criteria (read on demand, not all at once):

- {BIBLE_BUDDY}/references/anachronism-timeline.md → Read entire file (small, 43 lines) → for checks 3, 4 - {BIBLE_BUDDY}/references/commonly-misread-passages.md → Do NOT Read entire file (44KB wide table). Two-step lookup: 1. Load index: bash: grep '|' {BIBLE_BUDDY}/references/commonly-misread-passages.md | awk -F'|' '{print $2}' | sed 's/^ *//;s/ *$//' | grep -v '^-' | grep -v '^Scripture' → gives all 79 scripture references (~2KB) 2. For each relevant passage, Grep the full row: Grep("Isaiah 7:14", path="{BIBLE_BUDDY}/references/commonly-misread-passages.md") → for checks 2, 4, 7, 8 - {BIBLE_BUDDY}/references/yeshua-hermeneutics.md → Read entire file (small, 122 lines) → for check 8

1. Run checks sequentially:

- URL or pasted text: run checks 1-5 only - fun-facts reference file: run all 10 checks

1. For each check, report findings with line numbers, or "✅ 通過" if clean
2. End with a summary table

The 10 Checks

1. 事實錯誤檢查 (Factual Errors)

Reference files are the minimum, not the ceiling. Also apply your own biblical, historical, and linguistic knowledge to verify ALL factual claims — including date precision (e.g., "1820 年代" vs "1830 年代"), causal direction (did A cause B or vice versa?), and scholarly attribution accuracy.

Verify:

• Historical dates (cross-reference anachronism-timeline.md for known dates, AND verify precision — if timeline says "1820s-1830s" but content says only "1820 年代", flag the imprecision)
• Person attributions (who did what in which Bible chapter)
• Word meanings: if a Hebrew/Greek term appears in commonly-misread-passages.md's "Key Hebrew/Greek Term" column, verify the content uses it correctly. Also check semantic range — flag claims that reduce a word to a single meaning when it has a broader range (e.g., pistis ≠ only "faithfulness")
• Scholarly claims (hapax legomenon, manuscript evidence, etc.)
• If the content's interpretation of a passage contradicts the "Actual First-Century Jewish Context" column in commonly-misread-passages.md, flag it
• Causal claims: verify the direction is correct (e.g., if a Chinese translation was published in 1919, it cannot have "caused" an 1844 doctrine)
• Theological teachings without any scripture citation: if the content makes a theological claim (about God, the Spirit, salvation, church practice) without citing scripture, flag as "ungrounded teaching" (無經文根據的教導)

2. 經節編號檢查 (Verse Reference Errors)

Verify every chapter:verse reference matches the actual Bible content cited.

3. 年代錯置檢查 (Anachronism)

Cross-reference anachronism-timeline.md for every doctrine, practice, or concept mentioned. If a doctrine has a known origin date (e.g., altar call = 1830s, TULIP = 1618-1619, spiritual covering = 1970s), verify the content uses the correct date and does not project it back to the first century.

Also check date precision and conflation: if the timeline distinguishes between a doctrine (e.g., Dort's five points, 1618-1619) and a later label (e.g., TULIP acronym, 20th century), flag content that conflates them. If the timeline gives a range (e.g., "1820s-1830s"), flag content that uses only one end of the range without qualification.

Also detect implicit anachronistic terminology even when no verse is cited. Scan the content for terms that match any doctrine/practice in anachronism-timeline.md. If a match is found, note the actual origin date as supplementary context — not as a criticism. News articles and practical guides are not expected to cite historical origins. The goal is to give the reader background, not to fault the author for omitting academic footnotes.

Also flag entries that use post-200 CE sources (Mishnah, Talmud, Targums) to definitively explain pre-100 CE texts with language like "這才是...的背景". Later sources CAN illustrate earlier concepts, but the language must be hedged ("可能的背景" not "這才是"). Flag definitive claims only.

4. 過度簡化檢查 (Oversimplification)

Find entries that present one scholarly position as definitive when significant debate exists. Look for:

• Minority academic views stated as fact
• Complex debates reduced to one-sided conclusions
• Words like "其實是" or "就是" on debated topics

Cross-reference commonly-misread-passages.md: if the content falls into a "Common Misreading" pattern listed there, flag it. Also check yeshua-hermeneutics.md: if the content mentions Jesus' teaching methods (parables, arguments from Torah), verify it correctly identifies the method (e.g., kal va-chomer, mashal, remez) rather than oversimplifying as generic "metaphor" or "symbolism".

5. 數量計算檢查 (Counting Errors)

Verify all numerical claims: word/character counts, verse counts, people counts, ratios, dates.

6. 重複檢查 (Duplicates) ⟨fun-facts only⟩

Skip this check for URLs and pasted text.

Find entries that cover the same verse, topic, or argument. Two entries using the same framework applied to different topics are NOT duplicates. Two entries making the same point about the same verse ARE duplicates.

7. 稻草人檢查 (Strawman Arguments) ⟨fun-facts only⟩

Skip this check for URLs and pasted text.

Find entries where "不是 X" and X is something nobody in Taiwan would actually believe or confuse. Cross-reference commonly-misread-passages.md: if the denied X appears in the "Common Misreading" column, the denial is grounded and valid. If X has no basis in the reference file AND no basis in Taiwan church practice, it is likely a strawman. Test: does the misreading exist in the reference file, or would a Taiwan churchgoer actually think X?

8. 意味不明檢查 (Unclear Meaning) ⟨fun-facts only⟩

Skip this check for URLs and pasted text.

Find entries that just state a fact or quote a verse without explaining:

• What misconception it corrects, OR
• What surprising insight it reveals, OR
• Why a Taiwan reader should care

Every entry needs a clear "so what?"

If the entry involves a Hebrew/Greek term, check whether it provides the key term from commonly-misread-passages.md. Entries that discuss original language without naming the actual word (e.g., saying "原文意思是X" without giving the Hebrew/Greek) are weaker than those that do.

9. 字數控制檢查 (Verbosity) ⟨fun-facts only⟩

Skip this check for URLs and pasted text.

Compare each entry's character count to the median of the content. Flag entries significantly longer than the median (roughly 2x or more).

10. 趣味門檻檢查 (Interest Threshold) ⟨fun-facts only⟩

Skip this check for URLs and pasted text.

Flag entries that are too academic, too niche, or lack relevance for Taiwan readers. Test questions:

• Would a Taiwan Sunday school teacher find this interesting?
• Does it require specialized knowledge to appreciate?
• Is it relevant to what Taiwan churches actually teach or encounter?

Use commonly-misread-passages.md as a relevance guide: passages listed there with test cases from folk-religion, marketplace-theology, or charismatic-theology are highly relevant to Taiwan. Passages tied only to foreign contexts (e.g., Hindu theology, Vajrayana) are less relevant unless Taiwan has that community.

General Sweep (after completing all 10 checks)

Re-read the entire content one more time with fresh eyes. Use your own biblical, historical, and linguistic knowledge to catch anything the 10 structured checks missed. Common things to look for:

• Pre-existing antecedents that the content ignores (e.g., "Augustine invented original sin" ignores 4 Ezra, 2 Baruch)
• Conflation of related but distinct concepts (e.g., Colwell's rule vs Harner's qualitative argument)
• Translation history gaps (e.g., claiming CUV 1919 originated a connection when Morrison 1823 already used it)
• Counter-examples the content doesn't address (e.g., Joel 1:8 for betulah)

Report any additional findings under a separate ### 補充：通則檢查 section before the summary. If nothing additional is found, omit this section.

Output Format

Environment detection:

• Claude Code → uv run --directory {BIBLE_BUDDY} scripts/detect_desktop.py bible-fact-check → save to returned path
• Cowork / Claude.ai web → Do NOT save. Tell user: "你可以複製回應內容存檔，或在 Claude.ai 中使用 Artifact 功能下載。"

Filename: YYYYMMDD_HHmm_<slug>.md (URL → domain+path slug, pasted → pasted_text; slug 一律用 _ 分隔)

When reviewing a URL or pasted text (not a bible-buddy reference file):

1. Detect Desktop path using the script above
2. Write the full report to <detected-path>/YYYYMMDD_HHmm_<slug>.md
3. Show only the summary table in the conversation, with the file path

When reviewing a bible-buddy reference file: output the full report in the conversation as before.

Report structure

# 審查：[filename or URL or "使用者提供的內容"]
來源：[URL if applicable]
日期：[YYYY-MM-DD HH:mm:ss]

For each check that has findings, use a table:

### N. 檢查名稱

| # | 原文摘錄 | 問題說明 |
|---|---------|---------|
| 1 | 「...」 | 簡短說明 |
| 2 | 「...」 | 簡短說明 |

Checks with no findings: ### N. 檢查名稱 — ✅ 通過 (one line, no table)

End with a summary that explains the issues, not just counts them:

## 總結

| 檢查 | 結果 |
|------|------|
| 1. 事實錯誤 | 2 項 |
| 2. 經節編號 | ✅ |
| ... | ... |

### 主要問題說明

1. **[問題類型]** — 2-3 句白話解釋，這些問題為什麼重要、讀者該注意什麼。
2. ...

Readability Rules

Reports should be easy to read for a Taiwan churchgoer, not just scholars:

• 全中文為主 — 希臘文/希伯來文只在必要時附註，不要大段英文引用
• 短段落 — 每個發現 2-3 句話就好，不要寫論文
• 白話 — 用台灣教會能理解的語言，避免學術腔
• 補充背景 ≠ 批評 — 年代錯置檢查的語氣是「補充資訊」，不是指責作者
• Reference file 引用只需標示來源名稱和關鍵事實，不要貼原文

Important Notes

• Use 上主 (not 耶和華 or YHWH) when referring to the divine name in reports
• Do NOT auto-fix anything. Report only.
• Be specific: include line numbers and quote the problematic text
• When a finding is based on a reference file, cite it: 「根據 anachronism-timeline.md：altar call 起源為 1830s，非 1820s」
• When in doubt, flag it — the user will decide whether to act
• Some checks may not apply to all content types (e.g., key-terms lists won't have strawman arguments). Skip with "N/A — 此檢查不適用於本內容"
• Reference files: read ONLY the 3 files listed in Step 4 under {BIBLE_BUDDY}/references/ (resolved path).