Claude built a programming language for AI agents where every keyword is an internet meme
retained text
i had a dumb idea a while ago. what if ai agents had their own programming language, but every keyword was an internet meme? so i built **Memesis** with Claude Code. it's an actual interpreted language with its own `.meme` files, parser, runtime and CLI. **what it does** * `pepe` = an agent, `npc` = a function * `sus` = untrusted data * `yolo` = irreversible actions * `apu` = ask a human, `thonk` = call an ai model * `trustmebro` = mark untrusted data as reviewed * `chad / virgin / F` = try / catch / finally * `touchgrass` = sleep the funny part is that some of these actually make sense for ai agents. for example, data fetched from the internet gets marked as `sus`, and the runtime blocks it from reaching irreversible actions without an explicit trust check. there's also `memesis check` to catch issues before running anything and `memesis explain` to translate meme code into plain english. **how Claude Code helped** i came up with the concept and the meme keywords, then used Claude Code to implement the language itself. it wrote the lexer, parser, interpreter and CLI. the hardest part was the `sus` and `yolo` safety system. i described how untrusted data should be handled, and Claude Code implemented the tracking and static checks. it also wrote 34 tests and helped build the animated examples in the README using Playwright. it supports multiple agents, parallel tasks, Claude and local models through Ollama. **GitHub:** [https://github.com/tasdongyro/memesis](https://github.com/tasdongyro/memesis) it's completely free, open source under MIT and runs locally without signup. still experimental, but i'm curious what other Claude Code users think. what would you add or change? is there anything you'd approach differently?
document fetched 2026-11-06 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xook9e/claude_built_a_programming_language_for_ai_agents