KillSwitch

This page is a blog post by Dominik Picheta describing KillSwitch, an esoteric programming language built specifically to be difficult for large language models and coding agents to use. The author’s goal is to separate genuine code understanding from memorized snippets by testing models on a language they have never seen.

The language design deliberately targets LLM weaknesses. The spec contains hidden lies that humans are unlikely to read, and the language includes features that exploit difficulties with counting letters, over-eager safety guardrails, and the tendency to follow prompts too literally. The post walks through these design choices and then benchmarks major models against the language.

The content suits developers, researchers, and anyone interested in evaluating LLM coding ability beyond familiar benchmarks. Readers who want to try the language themselves or reproduce the benchmarks will find the design rationale useful, though the post is a write-up rather than a full documentation site.

Use it as a case study on benchmark contamination and prompt-following behavior rather than as a practical language reference. The benchmark results are worth reading alongside the design notes, since the two together explain why the language is hard for models.