Style Patterns

September 7, 2026 · View on GitHub

13. Em Dash Overuse

Long-form fire condition: 1+ em dash appears in each of 2+ consecutive paragraphs, or the document-level em-dash count is high relative to sentence count.

Short-form scoring branch (social / marketing register only): On English input of at most 200 non-whitespace characters and 1–4 prose sentences, count the em dashes that fall outside the exclusions below and score them directly: 1 → Low, 2 → Medium, 3+ → High. Always record em_dash_count and em_dash_per_sentence = em_dash_count / max(1, prose_sentence_count). This is a weak score/audit signal — one dash alone is not evidence the author is AI, and it is inert for the default profile. The deterministic engine (src/features/short-form.js) records it and the scorer floors an eligible reply off an exact 0 via computeShortFormEvidenceFloor in src/scoring.js (~1.7 for a single dash), which keeps it inside the human band rather than promoting it.

Exclusion: Do not trigger on interrupted speech in quoted dialogue, glossary or dictionary-style separators, or deliberate literary style where one or two parenthetical em dashes are clearly intentional.

Burstiness note: When rewriting em-dash-heavy passages, distribute the repairs across commas, sentence breaks, parentheses, and colons. Uniform substitution (replacing every em dash with a comma) produces a new rhythm problem.

Problem: AI overuses em dashes in place of commas, parentheses, or subordinate clauses to create a polished effect. Human writers use them sparingly across a document; one em dash per paragraph over several paragraphs is a tell.

Semantic Risk: LOW

Before:

The pilot launched in March — a date chosen to align with the grant cycle.

The first cohort included twelve clinics — each one serving a different rural county.

Early feedback was positive — though the team still needs retention data before expanding.

After:

The pilot launched in March, a date chosen to align with the grant cycle.

The first cohort included twelve clinics, each serving a different rural county.

Early feedback was positive. The team still needs retention data before expanding.

Short-form before / after (--profile social):

Before: built patina for exactly that — it keeps your meaning intact.

After: built patina for exactly that. it keeps your meaning intact.

The single dash reads as a light AI-polish tell in a promo reply; a period (or comma) keeps the same beat without it. This branch only scores/audits it — rewrite mode still leaves genuinely intentional single dashes alone.

14. Boldface Overuse

Fire condition: 5+ bolded terms in a single document, or 3+ in a single paragraph.

Exclusion: Bold is appropriate in UI documentation (button names, field labels), safety warnings ("Do not delete this file"), and genuine reference material where bold signals terminology. Only trigger in flowing prose.

Problem: AI bolds key terms as if writing a study guide or corporate slide deck. In normal prose, bolding should be rare or absent.

Semantic Risk: LOW

Before:

Machine learning is a subset of artificial intelligence that uses statistical methods to enable computers to learn from data. The most common approaches include supervised learning, unsupervised learning, and reinforcement learning.

After:

Machine learning is a subset of artificial intelligence that uses statistical methods to enable computers to learn from data. The most common approaches include supervised learning, unsupervised learning, and reinforcement learning.

15. Inline-Header Vertical Lists

Fire condition: 2+ "Label: explanation" bullets in the same list.

Exclusion: Legitimate reference content — API parameter tables, changelog entries, feature comparison grids — where label-and-description is the correct and expected format.

Burstiness note: When rewriting, avoid converting every bullet into a single run-on sentence. Vary the approach: some items can merge into prose, others can become numbered steps if order matters, and short items can stay as a plain list without bold labels.

Problem: AI converts prose into pseudo-heading bullet points with a bold label and colon on every item. This format is appropriate for reference docs but not for articles, essays, or narrative text.

Semantic Risk: LOW

Before:

  • Accessibility: The platform supports screen readers and keyboard navigation.
  • Performance: Load times have been reduced by 40%.
  • Security: All data is encrypted end-to-end.

After:

The platform supports screen readers and keyboard navigation. Load times have been reduced by 40%, and all data is encrypted end-to-end.

16. Title Case in Headings

Fire condition: 3+ content words (non-articles, non-prepositions) capitalized in the same heading when sentence case would be correct.

Exclusion: Proper nouns, brand names, product names, and acronyms within headings always require capitalization regardless of style guide.

Problem: AI capitalizes every word in headings (Title Case) instead of using sentence case. In most modern style guides (AP, APA, Google developer docs), sentence case is preferred for readability.

Semantic Risk: LOW

Before:

The Impact Of Remote Work On Urban Development And Housing Markets

After:

The impact of remote work on urban development and housing markets

17. Emojis

Watch words: emojis used as section markers or emphasis

Fire condition: Any emoji in professional, academic, or editorial text.

Exclusion: Personal blogs, social media copy, and casual newsletters where emojis are a deliberate stylistic choice. Also exclude emoji in quoted material being analyzed.

Problem: AI inserts emojis to add visual interest to text where they are inappropriate. This is especially common in listicles, summaries, and instructional content.

Semantic Risk: LOW

Before:

💤 Here are five tips for better sleep:

  1. Set a consistent bedtime
  2. Avoid screens before bed
  3. Keep your room cool
  4. Limit caffeine after noon
  5. Try reading instead of scrolling

After:

Here are five tips for better sleep:

  1. Set a consistent bedtime
  2. Avoid screens before bed
  3. Keep your room cool
  4. Limit caffeine after noon
  5. Try reading instead of scrolling

18. Curly Quotation Marks

Watch words: curly/smart quotes used in contexts where straight quotes are standard

Fire condition: Curly quotes appear inside code blocks, inline code spans, configuration file examples, or plain-text technical documents where straight quotes are required.

Exclusion: Curly quotes in narrative prose are typographically correct and not this pattern. Only trigger in technical or code contexts where curly quotes cause syntax errors or compatibility issues.

Problem: AI generates curly quotes in contexts where straight quotes are standard, such as code blocks, plain-text documents, configuration files, and technical documentation. This can cause syntax errors and compatibility issues.

Semantic Risk: LOW

Before:

Set the variable: name = “hello”

In your config file, add: timeout = ‘30’

After:

Set the variable: name = "hello"

In your config file, add: timeout = '30'

37. Colon Reveal / Plot-Twist Setup

Watch words: the best part:, the kicker:, the catch:, the result:, the twist:, plot twist:, the crazy part:, the wild part:, bonus:, spoiler:, the real reason:, translation:

Fire condition: A short noun phrase followed by a colon stages the rest of the sentence as a dramatic reveal, in prose (not a list, label, or definition). Fires on 2+ instances in a document, or 1 instance where a plain sentence loses nothing.

Exclusion: Colons doing structural work: list introductions, key–value labels, definitions, quote attributions, headlines and section labels, and code. A single colon reveal in intentionally punchy social copy the author wrote that way.

Semantic Risk: LOW

Problem: AI manufactures drama with a colon: a two-word setup, a beat, then the "reveal." Used once it is a rhythm choice; repeated, it turns every paragraph into a punchline delivery machine and reads as engagement-optimized copy. Rewrite as a plain sentence and keep the content.

Before:

The team shipped the migration in one weekend. The best part: zero downtime. The kicker: they did it with two engineers.

After:

The team shipped the migration in one weekend with zero downtime, using just two engineers.

Not this pattern: Em dashes used for the same beat --> Pattern 13 (Em Dash Overuse). Bold "Label: explanation" bullets --> Pattern 15 (Inline-Header Vertical Lists). Short-sentence stacking for rhythm --> en-viral-hook #4. Pattern 37 is specifically the colon-staged reveal inside prose.