Imported from Ryyyyyyy233/JF-Writing-Skill (jf-skill-1.0.0/SKILL.md). Install upstream with npx skills add Ryyyyyyy233/JF-Writing-Skill --skill jf-skill-1.0.0. Copyright stays with the author.

JF Writing Style Guide v3.1

Reference for writing and polishing academic finance prose in the Journal of Finance style. Extracted from 144 JF papers (2022--2025). Use for: drafting new sections, polishing drafts, Chinese-to-English translation.

v3.1 adds a three-tier AI-fingerprint defense (phrase scan, structural audit, quantitative verification), expanded anti-patterns covering structural symmetry, and a comprehensive detection-defense reference (Caveats M-R). Principles adapted from the structure-randomizer and deep-humanize skills.

Quick Start

Minimal invocation:

jf-skill 请帮我润色这段引言
JF风格 把这段中文摘要翻译成英文
jf-skill polish my results section

Execution flow:

1. Diagnose — Which section of the paper? (Introduction, Results, Conclusion, etc.)
2. Load references — Read the relevant reference file(s) for that section
3. Apply patterns — Match templates, calibrate hedging, check collocations
4. Polish — Iterate, avoiding mechanical substitution
5. Verify — Run AI-fingerprint check

Trigger Conditions

Trigger Keywords

中文: JF风格, JF润色, jf-skill, JF写作, JF翻译, JF句型, JF句式, JF用语

English: jf-skill, JF polish, JF translate, JF style, JF writing

When NOT to Use

Mode Selection

Reference Files Index

Core Workflow

Step 1: Diagnose the Context

Identify what section of the paper the text belongs to, what rhetorical function it serves, and what the user is asking for (polish existing, translate, draft new, restructure).

Step 2: Load Relevant References

Based on the diagnosis, read the appropriate reference files. Most tasks need only 1--2 references. A full paper audit needs all 5.

Step 3: Apply Patterns, Not Templates

The patterns in the references are descriptive (what JF authors actually do), not prescriptive (what you must do). Adapt them to the user's specific argument and evidence. Never copy-paste template sentences verbatim.

Step 4: Calibrate Hedging

Match claim strength to evidence strength. See the hedging calibration table in references/usage-guide.md.

• Speculative: "It is possible that..." / "We cannot rule out..."
• Suggestive: "These results suggest that..." / "This finding is consistent with..."
• Strong: "We find that..." / "Our results show that..."
• Definitive: "We conclude that..." / "Taken together, the evidence indicates that..."

Step 5: Three-Tier AI-Fingerprint Defense

AI text is statistically detectable: low structural variance, seamless transitions, zero admissions of uncertainty, flat information density. The correct defense is not lexical substitution but structural randomization — break symmetries, inject imperfection, vary every pattern that AI defaults to a mean. The three tiers below escalate from surface markers through structural diagnostics to quantitative thresholds.

Tier 1: Phrase-Level Scan (12 Markers)

Scan the full text. If any marker hits its threshold, fix mechanically.

Tier 2: Structural Self-Audit (5 Diagnostic Questions)

Read the full text and answer honestly. "Yes" to 3+ questions means structural rewrite is needed before submission.

1. How many distinct paragraph templates appear? Does any single template appear more than twice? If most paragraphs follow claim→evidence→implication, the paper has a structural AI fingerprint. At least 30% of paragraphs should deviate from the dominant template.

2. Is the Introduction the standard territory→gap→answer→roadmap four-segment structure? If yes: delete or compress the roadmap to one sentence. Consider merging territory and gap into one paragraph. Let the answer segment be 3-4 sentences, not a full preview of all findings.

3. Do parallel subsections (e.g., three Results subsections) use the same reporting template? If yes: vary their structures deliberately. One opens with the result; another opens with the intuition; a third opens with an unexpected finding. Make their lengths unequal — allow one subsection to be 1 paragraph while another is 3.

4. Are there at least 2 explicit admissions of uncertainty, surprise, or limitation? If no: add them. Human authors express doubt. Use "We do not yet have a clean explanation for..." or "This was not what we expected" or "Whether this holds beyond [boundary] we have not tested." Distribute uncertainty across different rhetorical types: data-level, mechanism-level, scope-level, and implicit (not using the word "uncertain").

5. Does the Conclusion recapitulate contributions one by one (numbered re-listing of what the Introduction already promised)? If yes: rewrite as 1 paragraph stating the core insight (1 sentence) + implication (1 sentence) + open question or concrete prediction (1 sentence). Never restate the contribution checklist.

Tier 3: Quantitative Verification (5 Metrics)

Compute these metrics on the final text. If any fails, return to the draft.

Metric calculation reference:

• Sentence length std dev: Extract all sentences, count words, compute std. If <25, deliberately inject 2-3 extremely short (≤8 word) sentences and 2-3 long (≥30 word) sentences into the densest section.
• Connector density: Grep for "however|therefore|thus|hence|moreover|furthermore|additionally|consequently|accordingly". Divide count by total word count × 1000.
• Extreme paragraph fraction: Count paragraphs with ≤3 sentences; count paragraphs with ≥8 sentences; sum and divide by total paragraph count.

Quality Standards

1. ⚠️ IRON RULE: Do not produce sentences a JF editor would flag as AI-generated. Run the fingerprint check every time.
2. ⚠️ IRON RULE: Match hedging to evidence strength. Overclaiming damages credibility more than underclaiming.
3. One paper, not all patterns. Overusing distinctive collocations ("shed new light on," "bridge different strands," "demystify") in the same paper creates artificiality. Each paper typically uses only 2--4 distinctive patterns.
4. Journal specificity. ALL CAPS openings, extensive enumeration, and roadmap paragraphs are JF conventions. Check target journal if not submitting to JF.
5. Contribution restraint. "To the best of our knowledge" requires a genuine literature search. "First direct evidence" is used by <5% of published papers.

Anti-Patterns

Caveats

Before every use, keep these in mind:

1. This is a reference, not a recipe. The patterns are what JF authors actually do; your paper has its own voice and argument. Use the skill to diagnose what's missing, not to replace your judgment.

2. Variation matters. A paper that uses every pattern from this skill will read as mechanical. Each real JF paper deploys only a subset.

3. Check your target journal. ALL CAPS openings, extensive enumeration, and "The remainder of the paper is organized as follows" are common in JF but may be inappropriate for JPE, QJE, or Econometrica.

4. The best JF writing is invisible. The goal is not to show off how many templates you've mastered — it's to make your argument clear, precise, and persuasive.

Version History

Synthesized from 144 Journal of Finance papers (2022--2025). All examples verbatim from published JF articles.