References
This page collects research papers, technical reports, and industry surveys related to cursor-rules design philosophy. These publications provide academic backing for the rationale of rule engineering.
Part 1: Prompt Engineering Foundations
[1] Brown et al. (2020) — Language Models are Few-Shot Learners
Source: Advances in Neural Information Processing Systems 33, OpenAI
Link: arxiv.org/abs/2005.14165
The GPT-3 paper systematically demonstrated for the first time that in-context learning can significantly change model output behavior without fine-tuning. This is the theoretical foundation of "rule injection" — providing constraints in prompts to guide models to follow specific norms.
"A few examples in the prompt can steer the model toward a desired behavior without gradient updates."
Relation to Rule Engineering: .mdc rule file injection is essentially an engineered application of in-context learning. Each AI interaction carries "few-shot examples" (rule conventions), guiding the model to generate code that conforms to project conventions.
BibTeX:
@inproceedings{brown2020language,
title={Language models are few-shot learners},
author={Brown, Tom and others},
booktitle={Advances in Neural Information Processing Systems},
volume={33},
pages={1877--1901},
year={2020}
}[2] Wei et al. (2022) — Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Source: NeurIPS 2022, Google Research
Link: arxiv.org/abs/2201.11903
Research shows that providing clear reasoning steps (rather than just answer examples) to LLMs can significantly improve code generation quality.
Relation to Rule Engineering: The "principle + example" structure of rule content is more effective than pure "what to do" checklists. Adding "why" and "how" explanations in rule files can improve AI compliance rates.
BibTeX:
@inproceedings{wei2022chain,
title={Chain-of-thought prompting elicits reasoning in large language models},
author={Wei, Jason and others},
booktitle={Advances in Neural Information Processing Systems},
volume={35},
pages={24824--24837},
year={2022}
}[3] Liu et al. (2023) — Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in NLP
Source: ACM Computing Surveys, arXiv:2107.13586
Link: arxiv.org/abs/2107.13586
A systematic survey of prompt engineering methods, covering hard prompt, soft prompt, instruction tuning and other technical routes.
Relation to Rule Engineering: .mdc rule files belong to the "hard prompt template" category — static, manually written instruction text. This survey provides a theoretical framework for understanding the advantages and disadvantages of this category.
Part 2: AI-Assisted Coding
[4] Copilot: GitHub Copilot Research Report (2022)
Source: GitHub / Microsoft Research
Link: github.blog
GitHub's research report shows that developers using Copilot complete tasks 55% faster, but code quality and consistency decline when context is missing.
Relation to Rule Engineering: The report implicitly explains the importance of context quality for AI-assisted coding effectiveness. Rule files are an engineering means to systematically improve context quality.
[5] Chen et al. (2021) — Evaluating Large Language Models Trained on Code (HumanEval)
Source: OpenAI
Link: arxiv.org/abs/2107.03374
The Codex model (GitHub Copilot's foundation) paper, introducing the HumanEval benchmark. The paper analyzes the "specification following" problem in code generation — whether models can accurately understand and follow natural language specifications.
Relation to Rule Engineering: Rule files are a lightweight "specification" mechanism, providing project conventions to models in natural language form. HumanEval's analysis framework can be used to evaluate the effectiveness of rule compliance.
BibTeX:
@article{chen2021evaluating,
title={Evaluating large language models trained on code},
author={Chen, Mark and others},
journal={arXiv preprint arXiv:2107.03374},
year={2021}
}[6] Jiang et al. (2023) — Self-planning Code Generation with Large Language Models
Source: arXiv:2303.06689
Link: arxiv.org/abs/2303.06689
Research on improving generation quality by letting LLMs plan before coding. Found that explicit structural constraints (like "design interface first, then implement details") can significantly improve code consistency.
Relation to Rule Engineering: Architecture conventions in rule files (like "define interfaces first", "use repository pattern") are essentially guiding the model's internal planning process.
Part 3: Software Engineering & AI Alignment
[7] Felten et al. (2023) — The Labor Market Impact of AI
Source: Princeton University, Princeton Policy Perspectives
Link: nber.org/papers/w31051
Analysis of AI's impact on software engineering careers, finding that AI is most efficient at standardized tasks, while tasks requiring deep context understanding still need human guidance.
Relation to Rule Engineering: Providing precise project context (i.e., rules) to AI can convert more "context-dependent" tasks into "standardized tasks", expanding AI's effective usage scope.
[8] Ouyang et al. (2022) — Training language models to follow instructions with human feedback (InstructGPT)
Source: NeurIPS 2022, OpenAI
Link: arxiv.org/abs/2203.02155
RLHF (Reinforcement Learning from Human Feedback) paper, showing that human-annotated preference data can significantly improve model instruction-following quality.
Relation to Rule Engineering: Models trained with RLHF (ChatGPT, Claude, GPT-4) have significantly better instruction-following capabilities than base models, making the directive writing style of rule files more effective.
BibTeX:
@inproceedings{ouyang2022training,
title={Training language models to follow instructions with human feedback},
author={Ouyang, Long and others},
booktitle={Advances in Neural Information Processing Systems},
volume={35},
pages={27730--27744},
year={2022}
}Part 4: Technical Articles & Industry Insights
[9] Simon Willison (2023) — Prompt Injection Attacks Against GPT-4
Link: simonwillison.net
Discussion of prompt injection attacks — security issues where malicious input overrides system-level instructions.
Relation to Rule Engineering: Understanding the security boundaries of rule injection. User code comments or string literals can theoretically interfere with rule injection effectiveness, a limitation rule engineering needs to be aware of.
[10] Andrej Karpathy (2015) — The Unreasonable Effectiveness of Recurrent Neural Networks
Link: karpathy.github.io
Although a 2015 article, its basic insight that "language models are statistical simulations of text" remains inspiring: model-generated code reflects the statistical properties of its training corpus.
Relation to Rule Engineering: Rule files, by introducing high-quality example text in prompts, effectively "guide" the model toward high-quality code distribution regions at inference time.
Further Reading
The following resources are not directly cited but highly relevant:
- Prompt Engineering Guide — Systematic prompt engineering knowledge base
- Anthropic's Prompt Engineering Documentation — Claude's official prompt engineering guide
- OpenAI Cookbook — Practice-oriented LLM application examples
- Continue.dev Documentation on Rules — Continue.dev rule system practice documentation
Export Citations
BibTeX Format
@misc{cursor-rules,
author = {LessUp},
title = {Cursor Rules: A Curated .mdc Rule Library for AI-Assisted Coding},
year = {2025},
publisher = {GitHub},
url = {https://github.com/LessUp/cursor-rules}
}
@inproceedings{brown2020language,
title={Language models are few-shot learners},
author={Brown, Tom and others},
booktitle={Advances in Neural Information Processing Systems},
volume={33},
pages={1877--1901},
year={2020}
}
@inproceedings{wei2022chain,
title={Chain-of-thought prompting elicits reasoning in large language models},
author={Wei, Jason and others},
booktitle={Advances in Neural Information Processing Systems},
volume={35},
pages={24824--24837},
year={2022}
}
@article{chen2021evaluating,
title={Evaluating large language models trained on code},
author={Chen, Mark and others},
journal={arXiv preprint arXiv:2107.03374},
year={2021}
}
@inproceedings{ouyang2022training,
title={Training language models to follow instructions with human feedback},
author={Ouyang, Long and others},
booktitle={Advances in Neural Information Processing Systems},
volume={35},
pages={27730--27744},
year={2022}
}Further Reading
- Related Work — Horizontal comparison of existing tool approaches
- Evolution — Future directions of rule engineering