Developers can chain design and implementation workflows in Claude Code
This repository connects the process of investigating a codebase, making design decisions, and writing actual code into a single command. Created by HumanLayer, this collection of skills is specialized for Claude Code, allowing complex development stages to be broken down and executed step by step. Each skill operates in an independent session, and specific stages can be invoked only when needed.
The rpi skill connecting design and implementation
The core feature is the rpi skill. This skill investigates changes, discusses design options with the user, and proceeds with implementation based on local documentation. You can choose from four paths: writing a PRD followed by TDD, technical design, or investigation followed by implementation. Each stage starts in a new session by default to prevent context pollution. The same functionality is supported when using Codex, requiring only an adjustment to file write permissions.
Auxiliary skills for visualization and code quality
In addition to rpi, several skills are included to assist the development process. show-me concisely explains the current topic using diagrams, code sketches, and HTML artifacts. visual-pr creates or updates pull requests with a visual overview so reviewers can easily understand the changes. improve-claude-md rewrites the CLAUDE.md file to increase instruction adherence, while narrow-react-prop-types narrows React component prop types to match actual code paths.
Tools for building agentic workflows
build-iterated-agentic-loop and design-control-loop are used to build more complex automation environments. The former generates repository-local skills, iterative coding agent GitHub Actions workflows, prompts, and memory files. The latter designs sensors, controllers, actuators, and disturbances through interviews, then implements them as locally executable components and scheduled coding agent workflows. This is the process of creating a custom control loop tailored to the codebase.
Installation and usage
All skills are installed using the npx command. For example, to install rpi, run npx skills add humanlayer/skills --skill rpi. After installation, invoke it in Claude Code with a slash command like /rpi, or use $rpi in Codex. Each skill is designed to focus on a specific task, allowing you to selectively execute only the necessary stages. When using Codex, you must grant write permissions to the .agents directory or specify a writable working directory.
Points of interest for Claude Code users
This repository is suitable for developers who use Claude Code as their primary tool. It is particularly useful for teams that want to separate design discussions from implementation or improve the code review process visually. For those working on React projects, the prop type cleanup skill may be helpful. However, since all skills are optimized for the Claude Code environment, compatibility should be verified if you primarily use other AI coding tools. Codex support is limited, and attention is required for file system permission settings.
