I built a programming language using Claude Code
Over the course of four weeks in January and February, I built a new programming language using Claude Code. I named it Cutlet after my cat. It’s completely legal to do that. You can find the source code on GitHub , along with build instructions and example programs . I’ve been using LLM-assisted programming since the original GitHub Copilot release in 2021, but so far I’ve limited my use of LLMs to generating boilerplate and making specific, targeted changes to my projects. While working on Cutlet, though, I allowed Claude to generate every single line of code. I didn’t even read any of the code. Instead, I built guardrails to make sure it worked correctly (more on that later). I’m surprised by the results of this experiment. Cutlet exists today. It builds and runs on both macOS and Linux. It can execute real programs. There might be bugs hiding deep in its internals, but they’re probably no worse than ones you’d find in any other four-week-old programming language in the world. I have Feelings™ about all of this and what it means for my profession, but I want to give you a tour of the language before I get up on my soapbox. If you want to follow along, build the Cutlet interpreter from source and drop into a REPL using . Arrays and strings work as you’d expect in any dynamic language. Variables are declared with the keyword. Variable names can include dashes. Same syntax rules as Raku . The only type of number (so far) is a double. Here’s something cool: the meta-operator turns any regular binary operator into a vectorized operation over an array. In the next line, we’re multiplying every element of by 1.8, then adding 32 to each element of the resulting array. The operator is a zip operation. It zips two arrays into a map. Output text using the built-in function. This function returns , which is Cutlet’s version of . The meta operator also works with comparisons. Here’s another cool bit: you can index into an array using an array of booleans. This is a filter operation. It picks the element indexes corresponding to and discards those that correspond to . Here’s a shorter way of writing that. Let’s print this out with a user-friendly message. The operator concatenates strings and arrays. The built-in turns things into strings. The meta-operator in the prefix position acts as a reduce operation. Let’s find the average temperature. adds all the temperatures, and the built-in finds the length of the array. Let’s print this out nicely, too. Functions are declared with . Everything in Cutlet is an expression, including functions and conditionals. The last value produced by an expression in a function becomes its return value. Your own functions can work with too. Let’s reduce the temperatures with our function to find the hottest temperature. Cutlet can do a lot more. It has all the usual features you’d expect from a dynamic language: loops, objects, prototypal inheritance, mixins, a mark-and-sweep garbage collector, and a friendly REPL. We don’t have file I/O yet, and some fundamental constructs like error handling are still missing, but we’re getting there! See TUTORIAL.md in the git repository for the full documentation. I’m a frontend engineer and (occasional) designer. I’ve tried using LLMs for building web applications, but I’ve always run into limitations. In my experience, Claude and friends are scary good at writing complex business logic, but fare poorly on any task that requires visual design skills. Turns out describing responsive layouts and animations in English is not easy. No amount of screenshots and wireframes can communicate fluid layouts and animations to an LLM. I’ve wasted hours fighting with Claude about layout issues it swore it had fixed, but which I could still see plainly with my leaky human eyes. I’ve also found these tools to excel at producing cookie-cutter interfaces they’ve seen before in publicly available repositories, but they fall off when I want to do anything novel. I often work with clients building complex data visualizations for niche domains, and LLMs have comprehensively failed to produce useful outputs on these projects. On the other hand, I’d seen people accomplish incredible things using LLMs in the last few months, and I wanted to replicate those experiments myself. But my previous experience with LLMs suggested that I had to pick my project carefully. A small, dynamic programming language met all my requirements. Finally, this was also an experiment to figure out how far I could push agentic engineering. Could I compress six months of work into a few weeks? Could I build something that was beyond my own ability to build? What would my day-to-day work life look like if I went all-in on LLM-driven programming? I wanted to answer all these questions. I went into this experiment with some skepticism. My previous attempts at building something entirely using Claude Code hadn’t worked out. But this attempt has not only been successful, but produced results beyond what I’d imagined possible. I don’t hold the belief that all software in the future will be written by LLMs. But I do believe there is a large subset that can be partially or mostly outsourced to these new tools. Building Cutlet taught me something important: using LLMs to produce code does not mean you forget everything you’ve learned about building software. Agentic engineering requires careful planning, skill, craftsmanship, and discipline, just like any software worth building before generative AI. The skills required to work with coding agents might look different from typing code line-by-line into an editor, but they’re still very much the same engineering skills we’ve been sharpening all our careers. There is a lot of work involved in getting good output from LLMs. Agentic engineering does not mean dumping vague instructions into a chat box and harvesting the code that comes out. I believe there are four main skills you have to learn today in order to work effectively with coding agents: Models and harnesses are changing rapidly, so figuring out which problems LLMs are good at solving requires developing your intuition, talking to your peers, and keeping your ear to the ground. However, if you don’t want to stay up-to-date with a rapidly-changing field—and I wouldn’t judge you for it, it’s crazy out there—here are two questions you can ask yourself to figure out if your problem is LLM-shaped: If the answer to either of those questions is “no”, throwing AI at the problem is unlikely to yield good results. If the answer to both of them is “yes”, then you might find success with agentic engineering. The good news is that the cost of figuring this out is the price of a Claude Code subscription and one sacrificial lamb on your team willing to spend a month trying it out on your codebase. LLMs work with natural language, so learning to communicate your ideas using words has become crucial. If you can’t explain your ideas in writing to your co-workers, you can’t work effectively with coding agents. You can get a lot out of Claude Code using simple, vague, overly general prompts. But when you do that, you’re outsourcing a lot of your thinking and decision-making to the robot. This is fine for throwaway projects, but you probably want to be more careful when you’re building something you will put into production and maintain for years. You want to feed coding agents precisely written specifications that capture as much of your problem space as possible. While working on Cutlet, I spent most of my time writing, generating, reading, and correcting spec documents . For me, this was a new experience. I primarily work with early-stage startups, so for most of my career, I’ve treated my code as the spec. Writing formal specifications was an alien experience. Thankfully, I could rely on Claude to help me write most of these specifications. I was only comfortable doing this because Cutlet was an experiment. On a project I wanted to stake my reputation on, I might take the agent out of the equation altogether and write the specs myself. This was my general workflow while making any change to Cutlet: This workflow front-loaded the cognitive effort of making any change to the language. All the thinking happened before a single line of code was written, which is something I almost never do. For me, programming involves organically discovering the shape of a problem as I’m working on it. However, I’ve found that working that way with LLMs is difficult. They’re great at making sweeping changes to your codebase, but terrible at quick, iterative, organic development workflows. Maybe my workflow will evolve as inference gets faster and models become better, but until then, this waterfall-style model works best. I find this to be the most interesting and fun part of working with coding agents. It’s a whole new class of problem to solve! The core principle is this: coding agents are computer programs, and therefore have a limited view of the world they exist in. Their only window into the problem you’re trying to solve is the directory of code they can access. This doesn’t give them enough agency or information to be able to do a good job. So, to help them thrive, you must give them that agency and information in the form of tools they can use to reach out into the wider world. What does this mean in practice? It looks different for different projects, but this is what I did for Cutlet: All these tools and abilities guaranteed that any updates to the code resulted in a project that at least compiled and executed. But more importantly, they increased the information and agency Claude had access to, making it more effective at discovering and debugging problems without my intervention. If I keep working on this project, my main focus will be to give my agents even more insight into the artifact they are building, even more debugging tools, even more freedom, and even more access to useful information. You will want to come up with your own tooling that works for your specific project. If you’re building a Django app, you might want to give the agent access to a staging database. If you’re building a React app, you might want to give it access to a headless browser. There’s no single answer that works for every project, and I bet people are going to come up with some very interesting tools that allow LLMs to observe the results of their work in the real world. Coding agents can sometimes be inefficient in how they use the tools you give them. For example, while working on this project, sometimes Claude would run a command, decide its output was too long to fit into the context window, and run it again with the output piped to . Other times it would run , forget to the output for errors, and run it a second time to capture the output. This would result in the same expensive checks running multiple times in the course of making a single edit. These mistakes slowed down the agentic loop significantly. I could fix some of these performance bottlenecks by editing or changing the output of a custom script. But there were some issues that required more effort to discover and fix. I quickly got into the habit of observing the agent at work, noticing sequences of commands that the agent repeated over and over again, and turning them into scripts for the agent to call instead. Many of the scripts in Cutlet’s directory came about this way. This was very manual, very not-fun work. I’m hoping this becomes more automated as time goes on. Maybe a future version of Claude Code could review its own tool calling outputs and suggest scripts you could write for it? Of course, the most fruitful optimization was to run Claude inside Docker with and access. By doing this, I took myself out of the agentic loop. After a plan file had been produced, I didn’t want to hang around babysitting agents and saying every time they wanted to run . As Cutlet evolved, the infrastructure I built for Claude also evolved. Eventually, I captured many of the workflows Claude naturally followed as scripts, slash commands, or instructions in . I also learned where the agent stumbled most, and preempted those mistakes by giving it better instructions or scripts to run. The infrastructure I built for Claude was also valuable for me, the human working on the project. The same scripts that helped Claude automate its work also helped me accomplish common tasks quickly. As the project grows, this infrastructure will keep evolving along with it. Models change all the time. So do project requirements and workflows. I look at all this project infrastructure as an organic thing that will keep changing as long as the project is active. Now that it’s possible for individual developers to accomplish so much in such little time, is software engineering as a career dead? My answer to this question is nope, not at all . Software engineering skills are just as valuable today as they were before language models got good. If I hadn’t taken a compilers course in college and worked through Crafting Interpreters , I wouldn’t have been able to build Cutlet. I still had to make technical decisions that I could only make because I had (some) domain knowledge and experience. Besides, I had to learn a bunch of new skills in order to effectively work on Cutlet. These new skills also required technical knowledge. A strange and new and different kind of technical knowledge, but technical knowledge nonetheless. Before working on this project, I was worried about whether I’d have a job five years from now. But today I’m convinced that the world will continue to have a need for software engineers in the future. Our jobs will transform—and some people might not enjoy the new jobs anymore—but there will still be plenty of work for us to do. Maybe we’ll have even more work to do than before, since LLMs allow us to build a lot more software a lot faster. And for those of us who never want to touch LLMs, there will be domains where LLMs never make any inroads. My friends who work on low-level multimedia systems have found less success using LLMs compared to those who build webapps. This is likely to be the case for many years to come. Eventually, those jobs will transform, too, but it will be a far slower shift. Is it fair to say that I built Cutlet? After all, Claude did most of the work. What was my contribution here besides writing the prompts? Moreover, this experiment only worked because Claude had access to multiple language runtimes and computer science books in its training data. Without the work done by hundreds of programmers, academics, and writers who have freely donated their work to the public, this project wouldn’t have been possible. So who really built Cutlet? I don’t have a good answer to that. I’m comfortable taking credit for the care and feeding of the coding agent as it went about generating tokens, but I don’t feel a sense of ownership over the code itself. I don’t consider this “my” work. It doesn’t feel right. Maybe my feelings will change in the future, but I don’t quite see how. Because of my reservations about who this code really belongs to, I haven’t added a license to Cutlet’s GitHub repository. Cutlet belongs to the collective consciousness of every programming language designer, implementer, and educator to have released their work on the internet. (Also, it’s worth noting that Cutlet almost certainly includes code from the Lua and Python interpreters. It referred to those languages all the time when we talked about language features. I’ve also seen a ton of code from Crafting Interpreters making its way into the codebase with my own two fleshy eyes.) I’d be remiss if I didn’t include a note on mental health in this already mammoth blog post. It’s easy to get addicted to agentic engineering tools. While working on this project, I often found myself at my computer at midnight going “just one more prompt”, as if I was playing the world’s most obscure game of Civilization . I’m embarrassed to admit that I often had Claude Code churning away in the background when guests were over at my place, when I stepped into the shower, or when I went off to lunch. There’s a heady feeling that comes from accomplishing so much in such little time. More addictive than that is the unpredictability and randomness inherent to these tools. If you throw a problem at Claude, you can never tell what it will come up with. It could one-shot a difficult problem you’ve been stuck on for weeks, or it could make a huge mess. Just like a slot machine, you can never tell what might happen. That creates a strong urge to try using it for everything all the time. And just like with slot machines, the house always wins. These days, I set limits for how long and how often I’m allowed to use Claude. As LLMs become widely available, we as a society will have to figure out the best way to use them without destroying our mental health. This is the part I’m not very optimistic about. We have comprehensively failed to regulate or limit our use of social media, and I’m willing to bet we’ll have a repeat of that scenario with LLMs. Now that we can produce large volumes of code very quickly, what can we do that we couldn’t do before? This is another question I’m not equipped to answer fully at the moment. That said, one area where I can see LLMs being immediately of use to me personally is the ability to experiment very quickly. It’s very easy for me to try out ten different features in Cutlet because I just have to spec them out and walk away from the computer. Failed experiments cost almost nothing. Even if I can’t use the code Claude generates, having working prototypes helps me validate ideas quickly and discard bad ones early. I’ve also been able to radically reduce my dependency on third-party libraries in my JavaScript and Python projects. I often use LLMs to generate small utility functions that previously required pulling in dependencies from NPM or PyPI. But honestly, these changes are small beans. I can’t predict the larger societal changes that will come about because of AI agents. All I can say is programming will look radically different in 2030 than it does in 2026. This project was a proof of concept to see how far I could push Claude Code. I’m currently looking for a new contract as a frontend engineer, so I probably won’t have the time to keep working on Cutlet. I also have a few more ideas for pushing agentic programming further, so I’m likely to prioritize those over continuing work on Cutlet. When the mood strikes me, I might still add small features now and then to the language. Now that I’ve removed myself from the development loop, it doesn’t take a lot of time and effort. I might even do Advent of Code using Cutlet in December! Of course, if you work at Anthropic and want to give me money so I can keep running this experiment, I’m available for contract work for the next 8 months :) For now, I’m closing the book on Cutlet and moving on to other projects (and cat). Thanks to Shruti Sunderraman for proofreading this post. Also thanks to Cutlet the cat for walking across the keyboard and deleting all my work three times today. I didn’t want to solve a particularly novel problem, but I wanted the ability to sometimes steer the LLM into interesting directions. I didn’t want to manually verify LLM-generated code. I wanted to give the LLM specifications, test cases, documentation, and sample outputs, and make it do all the difficult work of figuring out if it was doing the right thing. I wanted to give the agent a strong feedback loop so it could run autonomously. I don’t like MCPs. I didn’t want to deal with them. So anything that required connecting to a browser, taking screenshots, or talking to an API over the network was automatically disqualified. I wanted to use a boring language with as few external dependencies as possible. LLMs know how to build language implementations because their training data contains thousands of existing implementations, papers, and CS books. I was intrigued by the idea of creating a “remix” language by picking and choosing features I enjoy from various existing languages. I could write a bunch of small deterministic programs along with their expected outputs to test the implementation. I could even get Claude to write them for me, giving me a potentially infinite number of test cases to verify that the language was working correctly. Language implementations can be tested from the command line, with purely textual inputs and outputs. No need to take screenshots or videos or set up fragile MCPs. There’s no better feedback loop for an agent than “run and until there are no more errors”. C is as boring as it gets, and there are a large number of language implementations built in C. Understanding which problems can be solved effectively using LLMs, which ones need a human in the loop, and which ones should be handled entirely by humans. Communicating your intent clearly and defining criteria for success. Creating an environment in which the LLM can do its best work. Monitoring and optimizing the agentic loop so the agent can work efficiently. For the problem you want to solve, is it possible to define and verify success criteria in an automated fashion? Have other people solved this problem—or a similar one—before? In other words, is your problem likely to be in the training data for an LLM? First, I’d present the LLM with a new feature (e.g. loops) or refactor (e.g. moving from a tree-walking interpreter to a bytecode VM). Then I’d have a conversation with it about how the change would work in the context of Cutlet, how other languages implemented it, design considerations, ideas we could steal from interesting/niche languages, etc. Just a casual back-and-forth, the same way you might talk to a co-worker. After I had a good handle on what the feature or change would look like, I’d ask the LLM to give me an implementation plan broken down into small steps. I’d review the plan and go back and forth with the LLM to refine it. We’d explore various corner cases, footguns, gotchas, missing pieces, and improvements. When I was happy with the plan, I’d ask the LLM to write it out to a file that would go into a directory. Sometimes we’d end up with 3-4 plan files for a single feature. This was intentional. I needed the plans to be human-readable, and I needed each plan to be an atomic unit I could roll back if things didn’t work out. They also served as a history of the project’s evolution. You can find all the historical plan files in the Cutlet repository. I’d read and review the generated plan file, go back and forth again with the LLM to make changes to it, and commit it when everything looked good. Finally, I’d fire up a Docker container, run Claude with all permissions—including access—and ask it to implement my plan. Comprehensive test suite . My project instructions told Claude to write tests and make sure they failed before writing any new code. Alongside, I asked it to run tests after making significant code changes or merging any branches. Armed with a constantly growing test suite, Claude was able to quickly identify and fix any regressions it introduced into the codebase. The tests also served as documentation and specification. Sample inputs and outputs . These were my integration tests. I added a number of example programs to the Cutlet repository—most of them written by Claude itself—that not only serve as documentation for humans, but also as an end-to-end test suite. The project instructions told Claude to run all of them and verify their output after every code change. Linters, formatters, and static analysis tools . Cutlet uses and to ensure a baseline of code quality. Just like with tests, the project instructions asked the LLM to run these tools after every major code change. I noticed that would often produce diagnostics that would force Claude to rewrite parts of the code. If I had access to some of the more expensive static analysis tools (such as Coverity ), I would have added them to my development process too. Memory safety tools . I asked Claude to create a target that rebuilt the entire project and test suite with ASan and UBSan enabled (with LSan riding along via ASan), then ran every test under the instrumented build. The project instructions included running this check at the end of implementing a plan. This caught memory errors—use-after-free, buffer overflows, undefined behavior—that neither the tests nor the linter could find. Running these tests took time and greatly slowed down the agent, but they caught even more issues than . Symbol indexes . The agent had access to and for navigating the source code. I don’t know how useful this was, because I rarely ever saw it use them. Most of the time it would just the code for symbols. I might remove this in the future. Runtime introspection tools . Early in the project, I asked Claude to give Cutlet the ability to dump the token stream, AST, and bytecode for any piece of code to the standard output before executing it. This allowed the agent to quickly figure out if it had introduced errors into any part of the execution pipeline without having to navigate the source code or drop into a debugger. Pipeline tracing . I asked Claude to write a Python script that fed a Cutlet program through the interpreter with debug flags to capture the full compilation pipeline : the token stream, the AST, and the bytecode disassembly. It then mapped each token type, AST node, and opcode back to the exact source locations in the parser, compiler, and VM where they were handled. When an agent needed to add a new language feature, it could run the tracer on an example of a similar existing feature to see precisely which files and functions to touch. I was very proud of this machinery, but I never saw Claude make much use of it either. Running with every possible permission . I wanted the agent to work autonomously and have access to every debugging tool it might want to use. To do this, I ran it inside a Docker container with enabled and full access. I believe this is the only practical way to use coding agents on large projects. Answering permissions prompts is cognitively taxing when you have five agents working in parallel, and restricting their ability to do whatever they want makes them less effective at their job. We will need to figure out all sorts of safety issues that arise when you give LLMs the ability to take full control of a system, but on this project, I was willing to accept the risks that come with YOLO mode.
AI
JavaScript
Web Development
TypeScript
Svelte
DevOps
Php
CSS
Open Source
Rust
Programming
Python
HTML
Design
Security
JSON
Backend
Web
Gaming
Tutorial
C
Java
C++