kda
Kernel Design Agents (KDA) is a agent-centric workflow to write high-performance CUDA Kernels.
Kernel Design Agents (KDA) is an agent skill designed to enable coding agents to research, implement, verify, and iteratively optimize high-performance CUDA kernels. It provides structured workflow documentation, prompt templates, and repository-facing instructions that guide AI agents through the complete lifecycle of performance-sensitive GPU kernel development. The skill includes integration with KernelWiki and NCU profiling tools to help agents analyze, benchmark, and improve kernel implementations systematically.
Key Features
Use Cases
- 01Guiding AI agents to implement optimized CUDA kernels for machine learning workloads
- 02Automating iterative GPU kernel performance tuning and profiling analysis
- 03Researching and prototyping high-performance computing solutions with agent assistance
- 04Benchmarking and comparing multiple CUDA kernel implementation candidates
- 05Training coding agents on GPU programming best practices through structured workflows
- 06Building competition-grade kernel implementations for MLSys contests and challenges
Related Skills
View moresuperpowers
An agentic skills framework & software development methodology that works.
skills
Skills for Real Engineers. Straight from my .agents directory.
skills
Public repository for Agent Skills
ponytail
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
kda — FAQ
What is Kernel Design Agents (KDA)?+
KDA is an agent skill that provides structured workflows, prompts, and integration tools to help coding agents research, implement, verify, and optimize CUDA kernels. It guides agents through the complete lifecycle of performance-sensitive GPU programming tasks.
How do I install the KDA agent skill?+
Clone the repository with submodules, then symlink the skills directory to ~/.claude/skills/ so Claude Code can access the KernelWiki and ncu-report-skill components. You also need to install the humanize plugin through Claude Code's plugin marketplace.
Which AI clients work with KDA?+
KDA is designed for Claude Code and other coding agents that support skill-based workflows. The documentation references Claude-specific features like the plugin marketplace and skill directory structure.
What prerequisites does KDA require?+
You need Git to clone the repository with submodules, access to Claude Code or a compatible coding agent, and the ability to install the humanize plugin. For actual kernel development, you'll need CUDA toolkit and GPU hardware for profiling.
Is Kernel Design Agents free to use?+
Yes, the first-party KDA content is licensed under Creative Commons Attribution 4.0 and Apache 2.0. The included third-party submodules (KernelWiki and ncu-report-skill) are MIT-licensed but may contain embedded artifacts with their own upstream terms.
How does KDA help agents optimize CUDA kernels?+
KDA provides a minimal flow where agents define task contracts, create implementation plans, iterate in small verified steps, and record profiling evidence and benchmark results. The workflow integrates NCU profiling reports and KernelWiki references to guide performance improvements.
How do I install kda?+
Open the source repository on GitHub and follow its README. kda is a skill — MCP Agents Market links you directly to the official repo.
Is kda free?+
kda is an open-source project hosted on GitHub. Check the repository for its license and any usage requirements.