How to adopt a production-ready agent and rule set for building real products with Claude Code
If you use Claude Code daily and find that context runs out as sessions get longer, or you have to repeat the same instructions for recurring tasks, this repository offers a solution. It is a collection of agents, skills, hooks, commands, rules, and MCP settings refined over more than 10 months of actual product development by an Anthropic hackathon winner. Once installed, you can adopt a structure that automates your development workflow, allowing Claude Code to review code, write tests, and maintain memory across sessions.
A configuration built on 10 months of usage by a hackathon winner
This repository is not just a collection of configuration files; it is a production-grade setup validated while building the actual product zenith.chat. It includes guides on advanced topics such as token optimization, memory persistence, parallel processing, and subagent orchestration, helping you understand why the configuration is structured that way rather than just copying it. All scripts have been rewritten in Node.js to ensure identical operation on Windows, macOS, and Linux.
A structure where agents, skills, and hooks work together
The repository systematically categorizes and provides Claude Code’s extension elements: agents, skills, hooks, commands, and rules. Agents are specialized subagents that perform specific roles like code review, security analysis, and TDD guidance, handling tasks independently without polluting the main session’s context. Skills are reusable task definitions such as TDD workflows, security checklists, and continuous learning patterns, while hooks handle event-based automation, such as loading context at session start or saving state at session end. You can also execute complex workflows with a single line using slash commands like /tdd, /plan, and /code-review.
Automatic package manager detection and cross-platform support
One of the latest updates to this repository is full cross-platform support via Node.js-based scripts. It automatically detects the user’s preferred package manager among npm, pnpm, yarn, and bun by checking environment variables, project settings, and lock files in that order. You can specify your desired manager by setting the CLAUDE_PACKAGE_MANAGER environment variable or running the scripts/setup-package-manager.js script. This feature allows team members using different package managers to develop without conflicts using the same configuration files. Additionally, you can verify the proper operation of scripts and hooks through the test suite included in the tests/ directory.
Context window management and customization principles
The README strongly warns against activating all MCP servers simultaneously. A 200k context window can shrink to 70k due to tool overload, and it is recommended to keep active MCPs under 10 and active tools under 80 per project. Disabling unused servers via the disabledMcpServers setting is essential. Since this configuration is optimized for the author’s workflow, you should modify it to fit your stack and aggressively remove unused items. It is best to start with the parts you like and gradually add your own patterns.
Things to check before applying
This repository is released under the MIT license, allowing free use and modification. However, if you are not familiar with the Claude Code plugin system, you must choose between the plugin installation method (plugin marketplace add) and manual installation based on your environment. Manual installation requires copying files to the ~/.claude/ directory and adding hook settings to settings.json, and you must replace API key placeholders with actual values when configuring MCP. Ignoring context window management rules and activating all tools will actually degrade performance, so you must strictly follow the README’s recommendations. If your development style conflicts with the repository’s rules, modifying or deleting the relevant rule file is the normal way to use it.
