Building a Programming Language with Claude Code: The Cutlet Experiment

Building Cutlet with AI-Generated Code

Ankur Sethi created a new programming language called Cutlet using Claude Code over four weeks in January and February. Unlike typical LLM-assisted programming where AI helps with boilerplate or targeted changes, Sethi had Claude generate every single line of code without reading any of it himself. Instead, he focused on building guardrails and testing to ensure correctness.

The resulting language exists today, builds and runs on both macOS and Linux, and can execute real programs. While there may be bugs, Sethi notes they're probably no worse than any other four-week-old programming language.

Cutlet Language Features

Cutlet is a dynamic language with these key features:

• Variables declared with my keyword: my cities = ["Tokyo", "Paris", "New York", "London", "Sydney"]
• Variable names can include dashes (same syntax rules as Raku)
• Single numeric type: double
• Arrays and strings work as expected in dynamic languages
• @ meta-operator for vectorized operations: temps-c @* 1.8 multiplies each array element
• @: operator for zipping arrays into maps: cities @: temps-f creates {Tokyo: 82.4, Paris: 71.6, ...}
• say() function for output, returns nothing (Cutlet's null)
• Boolean array indexing for filtering: cities[temps-f @> 75] returns [Tokyo, New York]
• ++ operator concatenates strings and arrays
• str() built-in converts to strings
• Prefix @ for reduce operations: @+ temps-c sums all temperatures
• len() built-in finds array length
• Functions declared with fn: fn max(a, b) is ... if a > b then a else b ... end
• Everything is an expression, including functions and conditionals
• Custom functions work with @ operator: @max temps-c reduces with user-defined max function

Additional features include loops, objects, prototypal inheritance, mixins, mark-and-sweep garbage collector, and a friendly REPL. File I/O and error handling are not yet implemented.

Development Approach

Sethi built the interpreter from source and used /path/to/cutlet repl to drop into a REPL. The source code is available on GitHub with build instructions and example programs. He's been using LLM-assisted programming since GitHub Copilot's 2021 release but previously limited AI to boilerplate and targeted changes.

This experiment represents a shift to having the AI generate all code while the developer focuses on higher-level structure, testing, and guardrails. The approach proved surprisingly effective for creating a functional programming language.