Skip to content

how-to-use-with-claude-code.md

Latest commit

 

History

History
88 lines (56 loc) · 3.03 KB

how-to-use-with-claude-code.md

File metadata and controls

88 lines (56 loc) · 3.03 KB
Raw

How to Use kernel-skills with Claude Code

Claude Code is a terminal-based AI coding agent from Anthropic. You can paste a SKILL.md file directly into your request to guide Claude's reasoning process when writing, debugging, or optimizing compute kernels.

Basic usage pattern

  1. Open the SKILL.md file for the task you are working on.
  2. Copy its contents.
  3. Start a Claude Code session and paste the skill at the beginning of your request, followed by your specific task.

Example

<paste contents of skills/cuda/write-cuda-reduction-kernel/SKILL.md>

Using the skill above, write a warp-shuffle reduction kernel for the following:
- Input: float32, shape [1024, 2048]
- Operation: sum reduction along the last dimension
- Output: float32, shape [1024]
- Target: A100 (sm_80)
- Expected output dtype: float32

Claude Code will use the skill's reasoning process and design constraints to produce a more carefully considered kernel than it would from a bare prompt.

Using skills in ongoing sessions

For sessions where you are iterating on a kernel across multiple messages, include the skill in your first message. You do not need to repeat it in follow-up messages — Claude Code retains it in context.

Example: iterative debugging session

<paste contents of skills/cuda/debug-cuda-kernel-correctness/SKILL.md>

I have a softmax kernel that produces wrong results for input width 63 but correct results for width 64. Here is the kernel:

[paste kernel code]

Then follow up with specific questions:

Using the classification from the skill: is this a boundary condition bug or an indexing bug?

Combining multiple skills

For complex tasks it can be useful to provide two skills — for example, a "write" skill and a "correctness" skill together.

Example: write and verify

<paste contents of skills/triton/write-triton-attention-kernel/SKILL.md>
<paste contents of skills/patterns/write-kernel-test-plan/SKILL.md>

Write a causal attention kernel in Triton for:
- Q, K, V: float16, shape [batch=4, heads=8, seq=2048, head_dim=64]
- Causal mask: yes
- Target: H100 (sm_90)

After writing the kernel, also produce a test plan using the test plan skill.

Tips for best results

  • Provide all of the "Inputs the agent should gather first" fields listed in the skill. If you skip them, Claude Code will ask for them before writing code.
  • Specify hardware targets explicitly (e.g., "A100 sm_80", "H100 sm_90a"). Skills use this to make concrete tiling and type decisions.
  • If Claude Code produces code that violates a rule in the skill (e.g., hardcodes a tile size without justification), quote the specific rule back to it and ask for a revision.

Referencing skills from CLAUDE.md

If you are working in a repository that has a CLAUDE.md file (like this one), you can instruct Claude Code to reference specific skills by path:

Read skills/cuda/optimize-shared-memory-tiling/SKILL.md and apply it to the kernel in src/attention.cu

Claude Code will read the file automatically from the workspace.