LLM Writing Tropes Catalog: Community-Curated List of AI Text Patterns

LLM Writing Tropes (tropes.fyi) is a community-curated catalog documenting characteristic patterns and tell-tale phrases produced by large language models in generated text, covering stylistic quirks across GPT-4, Claude, Gemini, and Llama model families. The project has attracted 91 HN points and active developer discussion about AI text detectability, model fingerprinting, and the implications for AI-generated content in codebases and documentation. Engineers and AI developers working with LLM output quality or building AI detection tools will find the pattern list practically useful for calibrating or evaluating model outputs.

Key Takeaways

• tropes.fyi catalogs common LLM-generated text patterns including "certainly!", "I cannot assist with", hedge-stacking, and hallmarked structural templates across GPT-4, Claude, and Llama models
• Community-sourced with 91 HN points; covers text fingerprinting useful for developers building AI content detection, quality filters, or eval pipelines
• Relevant as AI-generated content proliferates in codebases, docs, and PRs — knowing model tropes helps distinguish AI from human-written contributions

Original source: tropes.fyi / Hacker News