Style Learner

Style Learning Skill

A style profile is metrics plus exemplars. Either alone is too weak to reproduce a voice.

Extract style from exemplar text and codify it as a profile that downstream skills (scribe:doc-generator, scribe:voice-generate) can apply consistently.

Approach: Feature Extraction and Exemplar Reference

The skill combines two methods because each fails alone:

1. Feature Extraction: quantifiable metrics (sentence length distribution, vocabulary complexity, structural patterns). Reproducible but soulless. 2. Exemplar Reference: specific passages that demonstrate the target style. Vivid but hard to apply at scale.

Together they form a profile precise enough to score new text and rich enough to guide rewrites. Metrics catch what exemplars miss. Exemplars carry what metrics flatten.

Required TodoWrite Items

1. style-learner:exemplar-collected - Source texts gathered 2. style-learner:features-extracted - Quantitative metrics computed 3. style-learner:exemplars-selected - Representative passages identified 4. style-learner:profile-generated - Style guide created 5. style-learner:validation-complete - Profile tested against new content

Step 1: Collect Exemplar Text

Gather representative samples of the target style.

Minimum requirements:

• At least 1000 words of exemplar text
• Multiple samples preferred (shows consistency)
• Same genre/context as target output

## Exemplar Sources

| Source | Word Count | Type |
|--------|------------|------|
| README.md | 850 | Technical |
| blog-post-1.md | 1200 | Narrative |
| api-guide.md | 2100 | Reference |

Step 2: Feature Extraction

Load: @modules/feature-extraction.md

Vocabulary Metrics

Sentence Metrics

Structural Metrics

Punctuation Profile

Step 3: Exemplar Selection

Load: @modules/exemplar-reference.md

Select 3-5 passages (50-150 words each) that best represent the target style.

Selection criteria:

• Demonstrates characteristic sentence rhythm
• Shows typical vocabulary choices
• Represents the desired tone
• Avoids atypical or exceptional passages

Exemplar Template

### Exemplar 1: [Label]
**Source**: [filename, lines X-Y]
**Demonstrates**: [what aspect of style]

> [Quoted passage]

**Key characteristics**:
- [Observation 1]
- [Observation 2]

Step 4: Generate Style Profile

Combine extracted features and exemplars into a usable style guide.

Profile Format

# Style Profile: [Name]
# Generated: [Date]
# Exemplar sources: [List]

voice:
  tone: [professional/casual/academic/conversational]
  perspective: [first-person/third-person/second-person]
  formality: [formal/neutral/informal]

vocabulary:
  average_word_length: X.X
  jargon_level: [none/light/moderate/heavy]
  contractions: [avoid/occasional/frequent]
  preferred_terms:
    - "use" over "utilize"
    - "help" over "facilitate"
  avoided_terms:
    - delve
    - leverage
    - comprehensive

sentences:
  average_length: XX words
  length_variance: [low/medium/high]
  fragments_allowed: [yes/no/sparingly]
  questions_used: [yes/no/sparingly]

structure:
  paragraphs: [short/medium/long] (X-Y sentences)
  lists: [prefer prose/balanced/prefer lists]
  headers: [descriptive/terse/question-style]

punctuation:
  em_dashes: [avoid/sparingly/freely]
  semicolons: [avoid/sparingly/freely]
  oxford_comma: [yes/no]

exemplars:
  - label: "[Exemplar 1 label]"
    text: |
      [Quoted passage]
  - label: "[Exemplar 2 label]"
    text: |
      [Quoted passage]

anti_patterns:
  - [Pattern to avoid 1]
  - [Pattern to avoid 2]

Step 5: Validation

Test the profile against new content:

1. Generate sample content using the profile 2. Compare metrics to extracted features 3. Have user evaluate voice/tone match 4. Refine profile based on feedback

Validation Checklist

• [ ] Metrics within 20% of exemplar averages
• [ ] No anti-pattern violations
• [ ] Tone matches user expectation
• [ ] Vocabulary aligns with exemplars
• [ ] Structure follows profile guidelines

Usage in Generation

When generating new content, reference the profile:

Generate [content type] following the style profile:
- Voice: [from profile]
- Sentence length: target ~[X] words, vary between [Y-Z]
- Use exemplar passage as tone reference:
  > [exemplar quote]
- Avoid: [anti-patterns from profile]

Module Reference

• See modules/style-application.md for applying learned styles to new content

Integration with slop-detector

After generating content, run slop-detector to verify: 1. No AI markers introduced 2. Style metrics match profile 3. Anti-patterns avoided

Exit Criteria

• Style profile document created
• At least 3 exemplar passages included
• Quantitative metrics extracted
• Anti-patterns from slop-detector integrated
• Validation test passed

Exemplar Reference Module

Select and document representative passages for style guidance.

Selection Criteria

Choose passages that demonstrate:

1. Characteristic rhythm: Sentence length variation patterns 2. Vocabulary choices: Typical word selection 3. Tone markers: How formality/informality is expressed 4. Structural preferences: Paragraph and list usage

Anti-Selection Criteria

Avoid passages that:

• Are atypically long or short
• Contain unusual formatting
• Quote external sources
• Are transitional/boilerplate
• Contain code blocks (unless style includes code)

Passage Length

Optimal exemplar length: 50-150 words

• Too short: Insufficient pattern demonstration
• Too long: Dilutes key characteristics

Annotation Format

### Exemplar: Technical Explanation

**Source**: docs/architecture.md, lines 45-52
**Word count**: 87
**Demonstrates**: Concise technical explanation with grounded examples

> The cache layer sits between the API and database. When a request
> arrives, we check Redis first. Cache hits return in under 5ms;
> misses fall through to Postgres, adding 50-200ms depending on
> query complexity. We chose Redis over Memcached for its richer
> data structures—sorted sets power our leaderboard feature.

**Key characteristics**:
- Short, direct sentences (avg 12 words)
- Specific numbers (5ms, 50-200ms)
- One em dash for aside
- Trade-off explanation ("chose X over Y because")
- No filler phrases
- Technical but accessible

Minimum Exemplar Set

For a complete style profile, collect at least:

For narrative content, add:

Usage in Generation

When generating new content, present exemplars as reference:

Write in a style similar to this passage:

> [exemplar text]

Key aspects to match:
- Sentence length around X words
- [Specific vocabulary preferences]
- [Tone markers to include]
- [Patterns to avoid]

Feature Extraction Module

Quantitative style metrics extraction from exemplar text.

Vocabulary Analysis

# Average word length
awk '{for(i=1;i<=NF;i++){sum+=length($i);count++}}END{print sum/count}' file.md

# Unique word ratio
words=$(tr '[:space:]' '\n' < file.md | grep -v '^$' | wc -l)
unique=$(tr '[:space:]' '\n' < file.md | grep -v '^$' | sort -u | wc -l)
echo "scale=2; $unique / $words" | bc

# Contraction count
grep -oE "\b\w+'(t|s|d|ll|ve|re|m)\b" file.md | wc -l

Sentence Analysis

import re

def analyze_sentences(text):
    # Split on sentence boundaries
    sentences = re.split(r'[.!?]+', text)
    sentences = [s.strip() for s in sentences if s.strip()]

    lengths = [len(s.split()) for s in sentences]

    return {
        'count': len(sentences),
        'avg_length': sum(lengths) / len(lengths),
        'min_length': min(lengths),
        'max_length': max(lengths),
        'std_dev': statistics.stdev(lengths) if len(lengths) > 1 else 0,
        'questions': sum(1 for s in sentences if '?' in s),
        'fragments': sum(1 for l in lengths if l < 5)
    }

Structural Analysis

# Paragraph lengths (sentences per paragraph)
awk -v RS='\n\n' '{
    gsub(/[.!?]/, "&\n");
    n = split($0, a, "\n");
    print n
}' file.md

# List ratio
bullets=$(grep -c '^\s*[-*]' file.md)
total=$(wc -l < file.md)
echo "scale=2; $bullets / $total" | bc

# Header depth
grep -E '^#{1,6}\s' file.md | head -1 | grep -o '#' | wc -c

Punctuation Profile

# Per 1000 words
words=$(wc -w < file.md)

em_dashes=$(grep -o '—' file.md | wc -l)
semicolons=$(grep -o ';' file.md | wc -l)
exclamations=$(grep -o '!' file.md | wc -l)
colons=$(grep -o ':' file.md | wc -l)

echo "Em dashes: $((em_dashes * 1000 / words)) per 1000"
echo "Semicolons: $((semicolons * 1000 / words)) per 1000"

Output Format

vocabulary:
  avg_word_length: 5.2
  unique_ratio: 0.42
  contraction_rate: 3.5  # per 100 sentences

sentences:
  avg_length: 18.4
  std_dev: 8.2
  question_rate: 2.1  # per 100
  fragment_rate: 1.5  # per 100

structure:
  avg_paragraph_sentences: 4.2
  list_ratio: 0.15
  max_header_depth: 3

punctuation:
  em_dash_rate: 1.8  # per 1000 words
  semicolon_rate: 0.5
  exclamation_rate: 0.2

Style Application Module

Apply learned style profiles to new content generation and editing.

Generation Prompting

When generating new content with a style profile:

## Style Guidelines

**Voice**: [profile.voice.tone] with [profile.voice.perspective] perspective

**Sentence targets**:
- Average length: [profile.sentences.average_length] words
- Vary between [min] and [max]
- [Fragment guidance from profile]

**Vocabulary**:
- Prefer: [profile.vocabulary.preferred_terms]
- Avoid: [profile.vocabulary.avoided_terms]
- Contractions: [profile.vocabulary.contractions]

**Structure**:
- Paragraphs: [profile.structure.paragraphs]
- Lists: [profile.structure.lists]

**Reference exemplar**:
> [Most relevant exemplar passage]

**Anti-patterns** (will be checked by slop-detector):
[profile.anti_patterns]

Editing to Match Style

When editing existing content to match a profile:

Step 1: Measure Current State

Extract metrics from current content and compare to profile.

Step 2: Prioritize Changes

1. High gap metrics first 2. Anti-pattern violations 3. Vocabulary substitutions 4. Structural adjustments

Step 3: Section-by-Section Editing

For each section: 1. Show current metrics 2. Propose specific changes 3. Present exemplar for reference 4. Wait for approval 5. Apply changes 6. Re-measure

Validation Loop

After applying style:

1. Run slop-detector on output
2. Re-extract metrics
3. Compare to profile targets
4. Flag remaining gaps > 20%
5. Iterate if needed

Style Drift Detection

For ongoing content:

# Compare new content metrics to profile
new_metrics=$(extract_metrics new-doc.md)
profile_metrics=$(cat .scribe/style-profile.yaml)

# Alert if drift > threshold
if [ $avg_sentence_diff -gt 5 ]; then
    echo "WARNING: Sentence length drifting from profile"
fi

Integration Points