I made Repomix: A Tool for Seamless Coding with Claude AI

TL;DR

I created Repomix, a Node.js tool that combines your project files into a single document. It was originally made to help me work with AI coding assistants like Claude, but it might be useful for other purposes too. It handles file encoding, ignores binary files, and respects .gitignore patterns.

I had a positive experience coding with Claude, which inspired me to create this Node.js command. It aims to simplify the process of working with AI on larger codebases. Here's how it works:

Running the following command in the root of your repository will create a file called repomix-output.txt:

$ npx repomix

You can then send this file to an AI assistant with a prompt like:

This file contains all the files in the repository combined into one. I want to refactor the code, so please review it first.

The AI can then understand the overall content and potentially assist with tasks like refactoring.

When you propose specific changes, the AI might be able to generate code accordingly. With features like Claude's Artifacts, you could potentially output multiple files, allowing for the generation of multiple interdependent pieces of code.

Features

This tool combines files under a specified folder into a single file.

https://github.com/yamadashy/repomix

While it handles details like ignoring .gitignore targets and binary files, and includes phrases to help AI understand, what it does is simple: it outputs a file like this:

================================================================
REPOMIX OUTPUT FILE
================================================================
// Phrases to help AI understand

================
File: src/index.js
================
// File contents

================
File: src/utils.js
================
// File contents

Claude 3.5 Sonnet can understand this simple structure well, so it can easily handle refactoring across multiple files.

While the context length that AI services can process is limited, this should be sufficient for small-scale projects. (Please don't use it for work code!)

You can adjust ignore targets in the configuration file. Please check the GitHub repository for more details.

Background

I've been experimenting with Claude Pro recently, using the Projects feature to load multiple files and have it modify code. It generates code with about 60-95% accuracy, which I find quite efficient when used as a reference for rewriting code.

However, the Projects feature doesn't allow uploading folders or zip files directly. I considered using Cursor, but hesitated due to its pay-as-you-go API.

Then I saw this article and realized that Claude could understand even when everything was in a single file, which led me to create and publish this tool.

https://qiita.com/kunishou/items/ed097b46cd78030e0b29

About the Development

While it's not doing anything too complex, I hope it can be of some use 🙏

Output File Content

The output file begins with text explaining how to handle the file.

Initially, I used a simple structure with just file paths and contents, but Claude sometimes had trouble understanding it. By adding "File Purpose," "Format," "Handling Method," and "Other Notes" at the beginning, Claude could understand more accurately.

Token Optimization

For file separators, I referred to ChatGPT's Tokenizer and used 16 and 64 = characters, which count as 1 token.

https://platform.openai.com/tokenizer

The token count might be different for ChatGPT-4 or Claude, but I aimed for something similar.

CLI Implementation

I used the commander library to create the command-line tool. I could have used yargs, but I like commander's simple syntax and automatic help generation.

To get the tool's version, I read from package.json. To ensure compatibility with older Node.js versions, I use fs to read it directly instead of import. Also, since __dirname isn't available in ES modules and import.meta.dirname isn't supported in some versions, I use a slightly tricky method:

import * as url from "url";

const dirName = url.fileURLToPath(new URL(".", import.meta.url));

Publishing as an npm Package

To easily publish new versions to npm, I added these scripts to package.json:

    "npm-publish": "npm run lint && npm run test-coverage && npm run build && npm publish",
    "npm-release-patch": "npm version patch && npm run npm-publish",
    "npm-release-minor": "npm version minor && npm run npm-publish",
    "npm-release-prerelease": "npm version prerelease && npm run npm-publish"

The npm version <major|minor|patch> command is surprisingly handy.

File Encoding

Since this tool shouldn't read binary files, I use is-binary-path to ignore binaries, and jschardet and iconv-lite to handle encoding properly.

File Ignoring

For ignoring specific files, the ignore package is useful. It makes it easy to implement .gitignore-like pattern filtering:

function createIgnoreFilter(patterns: string[]): (path: string) => boolean {
    const ig = ignore.add(patterns);
    return (filePath: string) => !ig.ignores(filePath);
}

Conclusion

This is a simple tool that I created because I needed it. While Claude will likely implement folder and zip upload features if we wait, I believe this tool can be used universally with the growing number of AI generation services.

I'd be happy to receive your feedback if you try it out!