What Happened
A developer named Ossama has published tropes.fyi, an open reference cataloging 40+ writing patterns that LLMs default to when generating text. The resource is organized into six categories and designed to be downloaded as a markdown file you can paste directly into system prompts.
The six categories cover word choice (5 tropes), sentence structure (8 tropes), paragraph structure (2 tropes), tone (11 tropes), formatting (3 tropes), and composition (7 tropes).
Some specific patterns called out:
- "Delve" and friends: The reflexive use of "certainly," "utilize," "leverage," "robust," "streamline," and "harness"
- Negative parallelism: The "It's not X - it's Y" construction that manufactures false profundity
- "Here's the kicker": Building false suspense before an unremarkable observation
- Invented concept labels: Fabricated terms like "supervision paradox" or "acceleration trap" that sound authoritative but mean nothing
- Em-dash addiction: Compulsive overuse of em dashes for dramatic pauses (20+ per document)
- Fractal summaries: Restating the same summary at every level of a document
- Bold-first bullets: Every list item starting with a bolded phrase
The resource emphasizes that individual tropes are not necessarily problems. The tell is clustering: when multiple patterns stack up together, the text reads as unmistakably AI-generated.
Why It Matters
If you use AI writing tools for content creation, marketing copy, blog posts, or documentation, you have run into these patterns. They are the reason AI-generated text often feels "off" even when it is technically correct. Readers have developed an instinct for detecting this kind of writing, and that instinct is getting sharper.
The practical value here is not just awareness. The markdown file format means you can add specific tropes to your system prompts or custom instructions in ChatGPT, Claude, or any other LLM. Instead of vaguely telling the model to "write naturally," you can point to exact patterns and say "do not do these things." That level of specificity produces noticeably better results.
For anyone publishing AI-assisted content professionally, this is a useful reference. Search engines are getting better at detecting AI-generated text, and readers are getting less tolerant of it. Eliminating these patterns is table stakes.
Our Take
This is one of those resources that makes you wince because you recognize every single pattern. We have caught ourselves using several of these tropes in our own AI-assisted drafts, including the "not X - it's Y" construction and the bold-first bullets.
The real insight in this catalog is the clustering observation. A single "delve" does not make text sound AI-generated. But "delve" plus negative parallelism plus em-dash abuse plus a fractal summary structure? That combination is a dead giveaway. The catalog trains you to spot these stacks, not just individual tics.
One caveat: if you dump all 40+ tropes into a system prompt, you will likely confuse the model or make it overly cautious. Pick the 5-10 patterns that bother you most and add those. Rotate them occasionally. The goal is targeted correction, not a comprehensive ban list that makes the LLM afraid to write anything.
We have been using a similar (shorter) list in our own content workflows for months. It works. The output still needs editing, but the starting point is noticeably less robotic. If you are doing any volume of AI-assisted writing, bookmark this one.