style-forensics（风格取证）

Style Forensics

Analyze: $ARGUMENTS

Audience: Writers and content creators who need a precise, measured voice fingerprint for use with style-writer and humanizer.

Goal: Extract quantitative style metrics from writing samples and produce a Style DNA report. Measure, don't guess.

Phase 1: Extract Prose

Strip the file to prose only before measuring anything. Remove:

• Code blocks (...)
• Inline code (replace with "CODE")
• Headings (lines starting with #)
• Blockquotes (lines starting with >)
• Horizontal rules (---)
• Metadata lines (tags, bylines, dates)
• URLs inside markdown links (keep link text)

Save the cleaned prose to a temp file for all subsequent analysis.

Report: Total prose word count.

Phase 2: Sentence Metrics

Extract individual sentences from the prose. Handle:

• Decimal numbers (don't split on "4.5")
• Abbreviations (e.g., i.e.)
• Quoted strings containing periods

Measure and report:

Describe the rhythm pattern: How do short and long sentences alternate? Do short sentences cluster? Are long sentences followed by punchy landings?

Phase 3: Punctuation DNA

Count across the full prose text:

Identify the punctuation fingerprint: Which marks carry the structural load? What's absent? Note any unusual ratios (e.g., colon-to-semicolon ratio).

Phase 4: Paragraph Structure

Phase 5: Vocabulary Profile

Describe the vocabulary character: Academic or accessible? Jargon-heavy or plain? Are key terms repeated deliberately (thematic anchoring) or is vocabulary varied?

Phase 6: Tone Profile

POV and Address

• Count I/my/me instances (per 100 words)
• Count we/our instances (per 100 words)
• Count you/your instances (per 100 words)
• Determine dominant POV

Hedging Language

Search for: probably, perhaps, maybe, might, could be, I think, kinda, sort of, likely, arguably, roughly, nearly, pretty, fairly, basically, genuinely, honestly, admittedly, I guess, I don't think

Report total, per 100 words, and each instance. Note whether hedges appear on claims, predictions, or data.

Emotional Register

• Profanity count and context for each instance
• Exclamation marks (already counted)
• Describe the overall temperature: restrained, casual, intense, breathless?

Contractions

• Total count and per 100 words
• Top 10 most frequent

Phase 7: Signature Devices

Search for and quantify each of these patterns. Not all will be present. Report what IS there and what IS NOT.

Sentence openers

• Top 15 opening words with counts
• Conjunction openers (And/But/Or/So): total count, percentage of all sentences, breakdown by word

Fragment patterns

• List all sentences under 6 words ending in a period
• Identify fragment stacking (2+ consecutive fragments)
• Identify single-word sentences

Recurring structures

Search for:

• "That's..." landing sentences
• "Here's..." launcher sentences
• "The [noun] is/was..." authoritative declaratives
• "Not X. Y." contrast patterns
• "I just..." casualness markers
• "Let me..." gear-shift cues
• "Same with X." stacking
• Credential references ("I've built/seen/written...")
• Direct persuasion ("believe me," "face it," "let's be honest")
• Any other repeated phrase pattern that appears 3+ times

Rhetorical devices

• Anaphora / repetition patterns
• Rhetorical question clusters (2+ questions in sequence)
• Colon setup-payoff structures
• Historical framing / analogy patterns

Temporal markers

• Count temporal words (first, then, next, eventually, now, started, etc.)
• Describe the temporal architecture: chronological? thematic? diary-style?

Phase 8: Structural Blueprint

Describe the macro-structure:

Opening pattern

How does the piece begin? (Bold thesis? Scene setting? Credential stack? Anecdote?)

Body organization

How are sections structured internally? What's the repeating unit?

Closing pattern

How does the piece end? (Call to action? Honest concession? Restatement? Promise?)

Phase 9: Compile the Report

Write the full Style DNA report as a markdown file saved next to the source file, named [source-filename]-style-dna.md.

The report MUST include:

1. Header with source file, word count, date analyzed
2. Measured Metrics table (all numbers from phases 2-4)
3. Sentence Length Distribution table
4. Vocabulary Profile table with top words
5. Tone Profile section (POV, hedging, emotional register)
6. Signature Devices section (only devices that actually appear, with counts and examples)
7. Structural Blueprint section
8. Quick Reference Card (code block with all key numbers for at-a-glance use)

If analyzing multiple files by the same author, add a Cross-Article Comparison section with a comparison table and analysis of which traits are stable fingerprints vs. sliding dials.

Execution Notes

• Run all independent measurements in parallel (sentence analysis, punctuation counts, vocabulary analysis can run simultaneously)
• Use Python for accurate counting. Shell one-liners are fine for simple counts but Python handles sentence splitting and edge cases better.
• Do not round aggressively. Report to 1 decimal place for percentages, 2 for per-100-word densities.
• List concrete examples for every device you identify. Numbers without examples are useless.
• If a metric reads zero, report it explicitly. Absence is data.

Phase 10: Save Prompt

After presenting the report, ask the user:

Style DNA report ready. Want me to save it as [source-filename]-style-dna.md next to the source file?

Wait for confirmation before writing. If the user says yes, save the full report. If analyzing multiple files, also offer to save the cross-article comparison as a separate file.

Integration

Produces Style DNA reports consumed by other skills:

[writing samples] --> style-forensics --> Style DNA report
                                              |
                                              +--> style-writer (write new content in that voice)
                                              +--> humanizer (preserve style when stripping AI tells)
                                              +--> copy-editor (respect voice during editing)

• Before style-forensics: Gather 1-5 writing samples (pre-AI preferred)
• After style-forensics: style-writer to write new content matching the voice
• Not this skill: brand-voice for qualitative brand guidelines, copy-editor for grammar review