humanizing-prose
Humanizing Prose
Avoid and fix the patterns that make text recognizable as AI-generated.
Two Modes
Writing Mode
When composing prose, internalize the tells catalog and avoid them from the start. Do not merely swap AI words for synonyms — write with varied rhythm, specific details, authentic voice, and no significance inflation.
Review Mode
When reviewing existing text, scan for tells systematically, then output a structured report:
## AI Tells Review
**Overall Assessment**: [Clean / Minor tells / Significant tells / Heavily AI-flavored]
**Vocabulary tells found**: [list with line references]
**Phrase tells found**: [list with line references]
**Structural tells**: [description]
**Tone issues**: [description]
### Suggested Rewrites
[For each flagged passage: original → suggested fix, with brief explanation]
Core Rules (When Writing)
- Vary sentence length — mix short punchy sentences with longer ones. Fragments are fine. Uniformity is the biggest structural tell.
- No hedging preambles — delete "It's important to note," "It's worth mentioning," etc. Just state the point.
- No significance inflation — don't call things pivotal, crucial, groundbreaking unless the user's content genuinely warrants it. Prefer concrete consequences over abstract importance claims.
- No participle analysis — never append "...highlighting its significance" or "...contributing to the broader ecosystem" to a factual statement.
- No negative parallelism — avoid "It's not X, it's Y" and "Not just X, but Y." These are the single most statistically common ChatGPT tell.
- No balanced-clause diplomacy — don't default to "While X, Y is also important." Take a position or state facts plainly.
- Specific > abstract — replace claims of impact with what actually happened. Numbers, names, consequences.
- Skip unnecessary transitions — "Additionally," "Furthermore," "Moreover," are AI tells. Just start the next sentence. Or use "And" / "But."
- Have a voice — allow opinions, asides, directness, even mild informality where appropriate. Monotone politeness is a tell.
- Don't over-structure — not everything needs headers and bullet points. Flowing paragraphs are often more natural.
- No em dashes — always replace em dashes (—) with alternative punctuation: commas, periods, colons, semicolons, or parentheses. Em dashes are a strong AI tell.
Reference Files
- Full tells catalog with categorized examples: references/tells-catalog.md
- Reformulation guide with before/after examples: references/reformulation-guide.md
Consult these references when doing a detailed review or when uncertain about a specific pattern.
Vocabulary Watchlist (Quick Reference)
Avoid or use sparingly: delve, crucial, pivotal, foster, bolster, underscore, highlight, emphasize, showcase, garner, intricate, meticulous, vibrant, enduring, profound, landscape (abstract), tapestry (abstract), testament, interplay, cornerstone, leverage, navigate, ecosystem, framework, groundbreaking, renowned, nestled, diverse array, boasts, seamlessly, notably, comprehensive
These aren't banned — context matters. "Pivotal" in a basketball article is fine. A paragraph with pivotal, crucial, fostering, and underscores is not.
More from markus1189/nixos-config
sourcegraph-search
Search code using Sourcegraph CLI. Use when (re)searching codebases, finding implementation examples, analyzing code patterns
22elfeed
Search the user's Emacs elfeed RSS feed database containing curated feeds from Reddit, blogs, YouTube, GitHub releases, and newsletters. Use when the user asks about articles they've read, mentions RSS feeds or 'something I read', wants to research topics from their curated sources (programming, AI, security, NixOS, Emacs, etc.), or needs to triage unread items.
1telegram
Direct integration with Telegram Bot API for checking inbox messages and sending messages (text, photos, documents) to chats. Triggers when users request 'check Telegram inbox', 'check messages', 'send message to [chat]', or mention Telegram bot interactions. Requires TELEGRAM_BOT_TOKEN environment variable.
1plantuml
Generates diagrams from text using PlantUML. Use when creating sequence diagrams, class diagrams, activity diagrams, C4 architecture diagrams, Gantt charts, mind maps, ER diagrams, or any visual diagram. Triggers on: diagram requests, architecture visualization, UML, flowcharts, system design, data modeling.
1transcribe-audio
Transcribes audio files (mp3, wav, ogg, m4a, flac, webm) using Gemini API via Portkey, saves transcripts as markdown, and supports follow-up analysis. Use when the user asks to transcribe audio, summarize a meeting recording, check a voice note, extract action items from a recording, asks what was discussed in an audio file, or mentions processing audio files in any way.
1nzb-search
Search and download NZB files from Usenet indexers (SceneNZBs, NZBgeek, NZBFinder, NZBPlanet) for movies, TV shows, books, and other media. Use when the user wants to find or download content from Usenet, mentions NZBs, asks for movies/TV/books with download intent, or wants to manage their cart.
1