Avoid AI Writing: A Deep Dive Into a 2,000-Line Rule System That Strips AI Flavor From Text
> Author: Conor Bronsdon > Repo: https://github.com/conorbronsdon/avoid-ai-writing > Version: v3.4.0, MIT License > Date: 2026-05-24
The post opens with the repo's standard demo—a paragraph stuffed with AI tells: chatbot openers ("Certainly!"), promotional filler ("vibrant", "nestled", "thriving"), meaning inflation ("watershed moment"), verb avoidance ("serves as", "featuring", "boasting"), empty appeals to authority ("experts believe"), transition filler ("Moreover"), and boilerplate endings ("the future looks bright").
The rewrite is one sentence:
> Acme Analytics raised a $40M Series B led by Sequoia. The Boulder-based startup makes an observability platform that runs queries in under a second and plugs into existing monitoring stacks without custom integration work.
Same information, but the second reads like a human wrote it. That is the tool's goal: pull AI's statistical average back to concrete human expression.
Three-Tier Vocabulary Gradient, Not a Blacklist
The rules classify vocabulary into three levels:
- Tier 1 — always replace. Words appearing 5–20x more often in AI text than human text: "tapestry", "realm", "paradigm", "beacon", "testament to", "game-changer", "watershed moment", "only time will tell" — nearly fifty terms total.
- Tier 2 — flag only in clusters. Words like "harness", "navigate", "foster", "elevate", "streamline", "empower" are fine alone; two or more in one paragraph trigger a rewrite.
- Tier 3 — flag by density. Words like "significant", "innovative", "compelling", "unprecedented", "sophisticated" are only flagged when overall density exceeds ~3%.
- Formatting: dash overuse, excessive bolding, heading emoji, bullet-point overload.
- Sentence patterns: "It's not X, it's Y" false contrasts, empty intensifiers ("genuinely", "truly"), vague endorsements, stacked hedging ("could potentially").
- Structure: metronomic paragraph lengths, grand-opening backgrounds, three-part symmetries, listicle inflation.
- Rhetoric: false concessions, rhetorical-question stalling, parenthetical hedges.
- Metadata leaks: chatbot politeness ("I hope this helps!"), tool URL tracking parameters, internal citation markers.
- Rhythm: structural regularity is the number-one detection signal. Even with all Tier 1 words replaced, uniform sentence length and templated transitions still read as AI. The Pangram classifier (trained on 28M human documents) ranks structural regularity above vocabulary.
- Stanford research (Liang et al., 2023): commercial AI detectors have >60% false-positive rates against non-native English writers.
- BFI working paper (Jabarian & Imas, 2025): open-source detectors misclassify >70% overall.
- arXiv:2506.07001 (2025): adversarial paraphrasing can reduce detection accuracy by ~88%.
- Conor Bronsdon, avoid-ai-writing (v3.4.0, MIT License): https://github.com/conorbronsdon/avoid-ai-writing
- Liang et al., "GPT detectors are biased against non-native English writers" (Patterns, 2023)
- Jabarian & Imas, BFI Working Paper 2025-116
- arXiv:2506.07001, adversarial paraphrase against AI detection
This turns detection into a probabilistic problem rather than a binary judgment.
Six Context Profiles
The same rules apply with different strictness per scenario:
| Context | Strictness | |---|---| | blog (default) | Full rules on | | technical-blog | Some Tier 2 terms relaxed | | investor-email | Extra strict on promotional language and meaning inflation | | linkedin | Structure relaxed, vocabulary strict | | docs | Relaxed overall, clarity first | | casual | Only P0 (credibility-killing) issues |
The insight: AI flavor is not absolute—it is deviation from context-appropriate expression.
A Morphology of 42 AI Tells
Beyond vocabulary, the skill catalogs 42 patterns:
The Most Underrated Rule: Don't Over-Polish
The final rule warns:
> Aggressively editing out every irregularity can push human writing *toward* AI statistical profiles. Natural disfluency, idiosyncratic word choices, and uneven pacing are what keep text out of the "AI-generated" classification.
AI output is statistically over-smoothed. Human writing stumbles, repeats, and jumps—features conventional editing removes, but which here are exactly the source of "human flavor".
Honest About Limitations
The skill opens with these figures:
The author's stance: "This is a writing-quality tool, not a verdict." A signal, not evidence—worth editing your own text by, not worth accusing others with.
Why AI Flavor Is Statistical Averaging
An LLM picks the most probable next token. Across billions of parameters, output converges toward an "average human"—flawless, never jarring, but stripped of any specific person's traits. "Leverage" wins over "use" because training data says so. The 2,000-line rule system's core job is pulling that statistical average back to the specific sentence, author, and context—hence the emphasis on second-pass audits, since rewrites can reintroduce averaged patterns in new packaging.
Conclusion
Released two months prior, the repo had gathered 1,400 stars and 142 forks despite being a single markdown file with no code or dependencies. The demand is real: people aren't looking for "better AI writing" but "writing that reads as human". The tool's real value is its systematic framework—how AI flavor can be identified across dimensions, how tolerance adjusts per context, and above all, that over-pursuing perfection is itself an AI tell.
> A signal, not proof. Worth editing by, not worth accusing with.
References