KernelWiki skill
Holds write-ups of GPU kernel optimizations for Blackwell and Hopper so an agent answers from sources, not model memory.
- 540 ★ on the repository
- Last commit
Minimal access
Why this level
- Scripts read base files inside the skill directory and use no network.
- Content is reference material: verify optimization advice with measurements.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/kernelwikiDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- Scripts read base files inside the skill directory and use no network.
- Content is reference material: verify optimization advice with measurements.
This is third-party code. Review the repository files before installing.
What it does
KernelWiki is a knowledge base on kernel optimization for NVIDIA Blackwell (SM100, B200) and Hopper (SM90, H100), packaged as a Claude Code skill. The repository root is the skill directory: reviewed pull requests, synthesized pages, blog and contest summaries, and indices by architecture, technique and hardware. The agent searches with `scripts/query.py`, `scripts/get_page.py` and `scripts/grep_wiki.py`, opens a page by id and follows it to the sources. Version-sensitive claims sit in a separate registry, and upstream code bundles are pinned to commit hashes through `PROVENANCE.yaml`.
Who it is for. Engineers writing CUDA, CUTLASS or Triton kernels for H100 and B200 who want agent answers backed by reviewed practice.
Good fit when
- An attention kernel runs slower than expected on B200 and you need techniques from recent pull requests.
- A team compares how projects use `tcgen05` instructions and MMA variants.
- An agent drafts an FP4 note and must cite concrete sources.
- You need the page for one technique fast: aliases like UMMA lead to the right term.
- Before moving from Hopper to Blackwell, the team reviews what changed in kernel scheduling.
Not a fit when
- The kernel targets AMD or Apple GPUs: the base covers NVIDIA only.
- You need information newer than May 2026: the pull request corpus closes on 20 May.
- The question is about general parallel computing theory, not specific kernels.
Example request
Look in KernelWiki for how FlashAttention-4 splits work between warps on B200 and show the sources.Limitations
The pull request corpus closes on 20 May 2026 and docs and leaderboards were verified through 18 August 2026. Blackwell and Hopper are covered, other architectures are not. The SSH clone address needs a GitHub key; clone over HTTPS without one. The Python scripts run without PyYAML but are faster with it. The repository is a submodule of Kernel Design Agents and feedback goes to the main repository. The content is reference material and does not replace measurements on your hardware.
How to disable. Delete `~/.claude/skills/KernelWiki` and restart Claude Code. Remove `BLACKWELL_WIKI_ROOT` from the shell profile if you set it.
FAQ
Do I need to install dependencies?
No. The scripts find the base root themselves and use a bundled PyYAML fallback. Install `requirements.txt` for speed.
How do I relocate the base?
Set `BLACKWELL_WIKI_ROOT` to the KernelWiki path and the scripts will read from it.
Where do claims come from?
Pages cite pull requests and docs. The `--follow-sources` option of `get_page.py` expands cited sources and `PROVENANCE.yaml` pins bundle versions.
Where do I report errors?
In the main Kernel Design Agents repository; this one is kept as a submodule for easy install.
Which architectures are covered?
Blackwell (SM100, B200) and Hopper (SM90, H100). Other architectures and vendors are not in the base.
Discussion
No questions or reviews yet.
Related
SkillsSkills for real engineers by Matt Pocock
Skills For Real Engineers
Small composable skills for engineering with agents: plan grilling, TDD, bug diagnosis, code review and architecture
A skills library that gives coding agents a development process: brainstorming, planning, TDD, subagents and code review
Karpathy-inspired Claude Code guidelines
Karpathy-Inspired Claude Code Guidelines
Four principles in one CLAUDE.md against common LLM coding mistakes: silent assumptions, overengineering, drive-by edits
Oh My Zsh is a shell configuration framework that keeps the old .zshrc separately so users can carry over functions and paths.
A plugin and skill that makes the agent write minimal code: reuse, stdlib and native platform features come first
GitHub toolkit for spec-driven development: the specify CLI adds agent commands and skills to a project, from principles to implementation
README in short
The README shows clone install, three query tools, a three-layer layout (sources, pages, indices) and quality gates run with `validate.py` and `verify_verbatim.py`.