Because I am not a nice person, I went off to find anti-patterns which would hurt LLM security analysis… turns out that there is indirect research which we can use.
Turns out that those nightmare macros in C which reference macros of macros going down seven layers are a nightmare for a human but also for LLMs.
Go FORTH and macro!
(sorry, I will see myself out)
For the curious the references I looked at are:
* Ernst, M.D., Badros, G.J., Notkin, D. "An Empirical Analysis of C Preprocessor Use." IEEE TSE, 2002, https://homes.cs.washington.edu/~mernst/pubs/c-preprocessor-tse2002.pdf
* arXiv:2504.15989v2. "Optimizing Token Consumption in LLMs: A Nano Surge Approach." 2025https://arxiv.org/html/2504.15989
* arXiv:2508.13666v1. "The Hidden Cost of Readability: How Code Formatting Silently Consumes Your LLM Budget." August 2025. https://arxiv.org/abs/2508.13666
* A. Airom, "The Token Economy of Programming Languages", Medium, August 2026 https://alain-airom.medium.com/the-token-economy-of-programming-languages-a1b59ee64513
* Tornhill, Adam. "Unhealthy Code is Burning Your Token Usage - Here's the Data."
CodeScene Blog, 28 May 2026.
https://codescene.com/blog/unhealthy-code-is-burning-your-token-usage-heres-the-data
* Cox, Landon & Kiefer, Mara. "Improving token efficiency in GitHub Agentic Workflows."
GitHub Blog, 12 May 2026. https://github.blog/ai-and-ml/github-copilot/improving-token-efficiency-in-github-agentic-workflows
* Anthropic. "Prompt Caching." Claude Platform Documentation.
https://platform.claude.com/docs/en/build-with-claude/prompt-caching
* pleasedodisturb. "Awesome LLM Token Optimization." GitHub repository.
https://github.com/pleasedodisturb/awesome-llm-token-optimization