Skip to content

how-to-use-with-gemini-cli.md

Latest commit

 

History

History
120 lines (78 loc) · 4.03 KB

how-to-use-with-gemini-cli.md

File metadata and controls

120 lines (78 loc) · 4.03 KB
Raw

How to Use kernel-skills with Gemini CLI

Gemini CLI is Google's terminal-based AI coding agent. It supports context injection via GEMINI.md files, direct file references, and inline prompting — all of which work well with kernel-skills SKILL.md files.

Basic usage pattern: paste into the prompt

The most direct approach: copy the SKILL.md content and include it in your Gemini CLI prompt.

Example

gemini "$(cat skills/cuda/write-cuda-reduction-kernel/SKILL.md)

Using the skill above, write a parallel warp-shuffle reduction kernel:
- Input: float32, shape [N=65536]
- Operation: sum reduction to a scalar
- Target: A100 (sm_80)
- Use warp shuffle intrinsics for the final warp reduction
- Provide a host-side launcher"

The $(cat ...) shell substitution inlines the skill content directly into the prompt string. Gemini CLI will respect the structured reasoning process in the skill.

Method 2: Use a GEMINI.md file for persistent context

Gemini CLI reads a GEMINI.md file from the current working directory (similar to Claude Code's CLAUDE.md). You can place skill content in GEMINI.md to apply it to all interactions in a project session.

Setup

# In your kernel project root:
cat skills/cuda/write-cuda-gemm-kernel/SKILL.md > GEMINI.md

Or combine multiple skills:

cat skills/cuda/write-cuda-gemm-kernel/SKILL.md \
    skills/cuda/optimize-shared-memory-tiling/SKILL.md \
    > GEMINI.md

Then launch Gemini CLI from that directory:

gemini

All interactions in that session will follow the skill instructions without needing to re-paste the content.

Method 3: Reference skill files directly

If kernel-skills is cloned inside your working directory or as a submodule, you can reference files directly in your prompt:

gemini "@skills/cuda/debug-cuda-kernel-correctness/SKILL.md

My kernel produces wrong results for inputs with N=2047. Here is the kernel:
$(cat src/my_reduction.cu)

Classify the bug using the skill and identify the root cause."

The @ file reference syntax tells Gemini CLI to read the file and include it in context.

Method 4: Multi-turn sessions with --interactive

For iterative kernel development, use Gemini CLI's interactive mode and include the skill at the start of the session:

gemini --interactive

In the first message:

I want you to follow this kernel engineering skill throughout our conversation:

[paste skill content]

Acknowledge the skill and tell me what inputs you need before writing any code.

Then continue the session with follow-up messages as you refine the kernel.

Example: optimization workflow

# Step 1: Write a first-draft kernel
gemini "$(cat skills/cuda/write-cuda-layernorm-kernel/SKILL.md)
Write a float16 LayerNorm kernel for shape [batch=64, hidden=4096], target sm_80."

# The output kernel is saved as layernorm_v1.cu

# Step 2: Optimize it
gemini "$(cat skills/cuda/optimize-shared-memory-tiling/SKILL.md)
Kernel to optimize:
$(cat layernorm_v1.cu)
Apply the shared memory tiling skill to identify memory access bottlenecks and propose improvements."

Tips for best results

  • For long skills, the $(cat file) substitution is more reliable than manual copy-paste since it avoids formatting loss.
  • Gemini CLI works well when the skill's "Inputs the agent should gather first" list is pre-filled. Provide all inputs (shapes, dtypes, hardware target) in the first message to avoid back-and-forth.
  • When asking Gemini to follow the review checklist: explicitly paste the checklist section at the end of your prompt and say "check each item and confirm compliance before finishing."
  • For multi-skill sessions, ordering matters: put the "write" skill before the "test plan" skill, or ask for one deliverable at a time per skill rather than requesting everything simultaneously.
  • Use a current strong reasoning or coding model. Kernel engineering tasks benefit significantly from a capable model — smaller or older versions produce noticeably weaker results for GPU kernel work.