AgentSkillsCN

prose-analysis

机械式散文指标——句长、开头多样性、对话比例、重复率、代词分布。当需要在做出主观判断前获得关于散文的量化信号，或在将草稿与项目基准对比时使用。

SKILL.md
--- frontmatter
name: prose-analysis
description: Mechanical prose metrics — sentence length, opener variety, dialogue ratio, repetition, pronoun distribution. Use when you need quantitative signals about prose before making subjective judgments, or when comparing a draft against the project's baseline.

Prose Analysis — Quantitative Signals

Scripts that measure mechanical properties of prose. These produce numbers, not quality verdicts. The numbers become useful only when compared against the project's own baseline — research shows no reliable universal thresholds for "good" prose exist.

What the Scripts Measure

The bundled resources/analyze.sh script produces:

  • •Sentence length distribution — mean, median, min, max, standard deviation. Compare across chapters for consistency. High variance often indicates varied rhythm (usually good); low variance may indicate monotone pacing.
  • •Sentence opener variety — categorizes first words of sentences (pronoun, article, conjunction, name, other). Heavy pronoun-leading ("I did... I saw... I felt...") is a common POV drift signal.
  • •Dialogue-to-narration ratio — proportion of lines that are dialogue vs narration. Useful for scene-type comparison (action scenes should skew narration, conversation scenes should skew dialogue).
  • •Repetition detection — words and short phrases that appear multiple times within a sliding window of N paragraphs. Catches echoed words the author didn't intend.
  • •Pronoun distribution — counts of I/me/my vs he/she/they/etc. Useful as a POV consistency check (a first-person chapter with sudden third-person pronouns needs investigation).

Running the Script

bash
bash resources/analyze.sh <file.md> [window_size]
  • •file.md — the markdown file to analyze
  • •window_size — optional, number of paragraphs for the repetition detection window (default: 5)

Output is plain text with labeled sections. Standard unix tools only (awk, sed, grep, wc, sort).

What the Scripts Don't Measure

Some metrics from the research literature are excluded because they're either unreliable, require Python NLP libraries, or don't produce actionable signals for creative writing:

  • •Readability scores (Flesch-Kincaid, etc.) — not a quality predictor for fiction
  • •Vocabulary richness (raw TTR) — methodologically weak at varying text lengths. MATTR is better but requires windowed computation that's impractical in pure bash.
  • •Perplexity — requires a language model to compute, and very low perplexity just means "predictable," not "bad"
  • •Adverb/filter word density — modest effect sizes, heavily genre-confounded

Resources

  • •resources/analyze.sh — the analysis script
  • •resources/antipatterns.md — AI writing antipatterns, honestly categorized as research-backed vs community folklore. Read this when reviewing AI-generated drafts and you want to know which signals are worth investigating vs which are unreliable.
  • •resources/baseline.md — how to establish a project baseline from existing chapters and compare new drafts against it. Read this before interpreting analysis output for the first time on a project.