A prompt you paste into any LLM to make AI-generated text sound like a person wrote it. The rules are grounded in peer-reviewed research on how AI text differs from human text, extended with patterns found by auditing current-model output.
PROMPT.md is a standalone document. You copy it into a conversation with Claude, ChatGPT, Gemini, Llama, or whatever you use. Paste your text at the bottom, hit send. The model edits your writing based on evidence about how AI text differs from human text at the word, sentence, tone, and cognitive levels.
This is not a detector. It does not classify text as human or AI.
It is an editing guide. The goal is writing that reads better: more varied and specific, less like it came off an assembly line. If your institution or jurisdiction requires you to disclose AI involvement, use this tool alongside that disclosure.
- Open any LLM chat interface.
- Copy the full contents of
PROMPT.md. - Paste it into the chat.
- Replace
[PASTE YOUR TEXT HERE]at the bottom with your text. - Send.
You get back an edited version with AI-correlated patterns reduced.
| Layer | What it addresses | Examples |
|---|---|---|
| Vocabulary | AI-overused words, nominalization, abstract nouns, adverb gaps, uniform hedging, credibility insistence | "delve" becomes "dig into"; "the implementation of" becomes "we implemented"; three uses of "real" become one |
| Structure | Sentence length variance, the antithesis reflex, section-to-section variation, syntactic rigidity, too many lists, dependency distance, punctuation variety | "It's not X, it's Y" becomes a direct claim; style shifts between opening, middle, and close; no three-item list stacking |
| Tone | Flattened sentiment, readability mismatch, narrow emotional range, missing negative affect | Allow frustration and skepticism; match grade level to the audience |
| Discourse | Formulaic transitions, over-explicit cohesion, coordination vs. subordination | Drop "Furthermore"; use "because" instead of "and" |
| Texture | Cognitive load artifacts, self-monitoring traces, abstraction-as-agent, domain vocabulary | Self-corrections; "the order mattered" gets a human subject |
The rules draw on the following peer-reviewed papers, reference documents, and community resources. Citations follow APA 7th edition format.
Ardeshirifar, S. (2025). Comparing handcrafted and deep learning approaches for detecting AI-generated text: Performance, generalization, and linguistic insights. AI and Ethics, 5, 4197–4209. https://doi.org/10.1007/s43681-025-00700-2
Kobak, D., et al. (2025). Delving into ChatGPT usage in academic writing through excess vocabulary. Science Advances.
Krishna, K., Song, Y., Karpinska, M., Wieting, J., & Iyyer, M. (2023). Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense. In Advances in Neural Information Processing Systems (NeurIPS 36).
Kuznetsov, K., et al. (2025). arXiv preprint, arXiv:2501.19301. [On the flow-level and section-to-section features that still separate AI from human text after models match human surface statistics.]
Muñoz-Ortiz, A., Gómez-Rodríguez, C., & Vilares, D. (2024). Contrasting linguistic patterns in human and LLM-generated news text. Artificial Intelligence Review, 57(10), 265. https://doi.org/10.1007/s10462-024-10903-2
O'Sullivan, J. (2025). Stylometric comparisons of human versus AI-generated creative writing. Humanities and Social Sciences Communications, 12, 1708. https://doi.org/10.1057/s41599-025-05986-3
Opara, C. (2025). Distinguishing AI-generated and human-written text through psycholinguistic analysis. arXiv preprint, arXiv:2505.01800. https://doi.org/10.48550/arXiv.2505.01800
Przystalski, K., Argasiński, J. K., Grabska-Gradzińska, I., & Ochab, J. K. (2024). Stylometry recognizes human and LLM-generated texts in short samples. Expert Systems with Applications, 296, 129001. https://doi.org/10.1016/j.eswa.2025.129001
Rujeedawa, T., et al. (2025). Unmasking AI-generated texts using linguistic and stylistic features. [Preprint].
Terčon, L., & Dobrovoljc, K. (2025). Linguistic characteristics of AI-generated text: A survey. [Preprint].
Tulchinskii, E., Kuznetsov, K., Kushnareva, L., Cherniavskii, D., Barannikov, S., Piontkovskaya, I., Nikolenko, S., & Burnaev, E. (2023). Intrinsic dimension estimation for robust detection of AI-generated texts. arXiv preprint, arXiv:2306.04723. https://doi.org/10.48550/arXiv.2306.04723
Grammarly. (2025). Common words and phrases in AI-generated text. https://www.grammarly.com/blog/ai-writing/common-ai-words/
Wikipedia contributors. (2025). Wikipedia:Signs of AI writing. In Wikipedia, The Free Encyclopedia. Retrieved April 13, 2026, from https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
The psycholinguistic layer builds on Cognitive Load Theory (Sweller, 1988), metacognition and self-monitoring (Flavell, 1979), lexical access models (Levelt, 1989), and discourse planning (Grabe & Kaplan, 1996). Opara (2025) maps stylometric features to these cognitive processes; that mapping shaped how the texture rules work.
Peer-reviewed and preprint studies from 2023 to 2025 provided the quantitative foundation. They cover lexical diversity, part-of-speech distribution, syntactic complexity, dependency structure, sentiment, n-gram patterns, and stylometric clustering.
A second stream comes from auditing current-model output directly. Some tells (the two-sentence antithesis flip, reality-insistence density) show up in frontier-model writing that passes the word-level checks clean. Where these overlap with the research, such as the contrastive-pattern work of Muñoz-Ortiz et al. (2024), the citation is noted; where they are observational, the prompt treats them as candidate rules rather than settled findings.
Findings were sorted into six layers: vocabulary, structure, tone, discourse, texture, and anti-patterns. Patterns confirmed by more than one independent research group got priority.
For each confirmed pattern, we wrote an actionable rewriting rule. Each rule says what to look for, how to change it, and, critically, how to apply the change unevenly. Uniform "humanizing" is itself an AI signal.
The anti-pattern examples show concrete before-and-after transformations. No abstract advice. The caveats section spells out where the rules break down: ESL contexts, genre differences, model evolution, and the paradox of consistent inconsistency.
Lexical diversity. AI text has lower Type-Token Ratios, fewer hapax legomena, fewer unique words (Ardeshirifar, 2025; Muñoz-Ortiz et al., 2024). The vocabulary diversity rules exist because of this.
Part-of-speech distribution. AI overuses auxiliary verbs, determiners, coordinating conjunctions, symbols, and numbers. Humans lean harder on nouns, adjectives, adverbs, and punctuation (Muñoz-Ortiz et al., 2024; Ardeshirifar, 2025). The gap is surprisingly consistent across datasets.
Human writers show wider sentence-length distributions with higher standard deviation; AI clusters around shorter, uniform lengths (Ardeshirifar, 2025; Terčon & Dobrovoljc, 2025). That variance is what the burstiness rules try to recreate.
Dependency length. Humans naturally keep syntactically linked words close together. LLMs are worse at this (Muñoz-Ortiz et al., 2024).
AI skews neutral-to-positive and suppresses negative emotions like anger and disgust (Muñoz-Ortiz et al., 2024). The tone rules push back against that flattening.
Stylometric clustering. In creative writing, LLM outputs form tight, uniform clusters while human texts scatter broadly, as measured by Burrows' Delta (O'Sullivan, 2025). If there is one finding that justifies the whole prompt, it is this one: stylistic individuality matters more than any single rule.
Flow and section-to-section variation. As models improve, they close the gap on word-order-independent surface statistics such as part-of-speech distribution and readability, showing little measurable difference from human writing. Two families of feature still diverge: the flow of the text (token-level unpredictability and how content shifts across it) and cross-segment variation, since humans modulate their style between a document's opening, body, and close while a model holds one flat fingerprint throughout (Kuznetsov et al., 2025). The arms race points the same way: paraphrasing evades most detectors (Krishna et al., 2023), and reliable evasion works by destroying exactly these flow features. The practical reading is that word substitution has a low ceiling, so the structural, discourse, and section-to-section rules carry more weight for frontier-model output.
AI uses fewer but more repetitive discourse markers, and fewer modal or epistemic markers (Terčon & Dobrovoljc, 2025). Nominalization rates run higher too (Terčon & Dobrovoljc, 2025), which is why the prompt pushes verb-led constructions.
Readability. AI writes at a higher grade level than humans do for the same content type (Terčon & Dobrovoljc, 2025; Opara, 2025). The calibration rules address that mismatch.
The flagged vocabulary reflects models through 2025: GPT-3.5, GPT-4, LLaMA, Falcon, Mistral. Newer models will adopt different word preferences. Structural and psycholinguistic rules should hold up longer than lexical ones.
Detection features break down under domain shift. Rules tuned for blog posts may misfire on academic prose. Ardeshirifar (2025) found a 34.67 percentage point F1 drop in cross-dataset testing.
Newer, larger models are converging on human-like word distributions, but structural uniformity still gives them away (O'Sullivan, 2025; Przystalski et al., 2024). Weight structural and psycholinguistic rules higher than word substitution when editing frontier model output.
Some patterns flagged here as AI-correlated are normal in certain human contexts, particularly academic and journalistic writing. Formulaic transitions and neutral tone are conventions in those genres. Keep that in mind before over-editing.
These rules describe statistical tendencies. Text that follows every rule can still be flagged. The goal is to shift the probability distribution toward human-like patterns.
Found a new paper on AI text detection patterns? Open an issue with the citation and a summary. Spotted a pattern the rules miss? Submit a pull request with the proposed rule and its research basis.
See CONTRIBUTING.md for details.
Released under the MIT Licence. The underlying research belongs to its respective authors and is cited above.