A curated collection of Claude Code agent skills that accelerate the entire vLLM development lifecycle — from understanding codebases and triaging issues to designing features, contributing changes, and analyzing performance. Each skill encapsulates a specific development task as a reusable, composable workflow, enabling vLLM developers and community contributors to work more efficiently with AI-assisted tooling.
| Name | Category | Description | Rec |
|---|---|---|---|
| vllm-dev-task-discovery | Demand Discovery | Analyze contribution opportunities in the vLLM repository for community developers. | ⭐️⭐️⭐️ |
| vllm-feature-tutorial | Code Reading | Generate comprehensive Chinese technical tutorial documents for vLLM features and modules. | ⭐️⭐️⭐️⭐️⭐️ |
| vllm-model-tutorial | Code Reading | Generate comprehensive Chinese technical tutorial documents for vLLM models. | ⭐️⭐️⭐️⭐️⭐️ |
| vllm-feature-design | Change Proposing | Design and implement vLLM features according to user requirements. | ⭐️⭐️⭐️⭐️ |
| vllm-rfc-generator | Change Proposing | Generate a vLLM-style RFC document based on user input. | ⭐️ |
| vllm-pr-desc-generator | Code Contribution | Generate a vLLM-style PR description from a GitHub PR's code changes. | ⭐️⭐️⭐️⭐️ |
| vllm-test-generator | Code Contribution | Generate unit tests or end-to-end tests for vllm. | ⭐️⭐️⭐️ |
| vllm-benchmark-result-analysis | Performance Optimization | Compare serving benchmark outputs before and after a code change. | ⭐️⭐️ |
| vllm-vs-atom-decode-trace-comparison | Performance Optimization | Compare decode-phase kernel implementations of vLLM vs ATOM from torch-profiler traces, with per-layer-type comparison tables and a vLLM optimization TODO list. | ⭐️⭐️⭐️⭐️ |
| vllm-pr-summary | Code Review | Fetch and analyze a PR from vllm, then generate a report covering PR overview, code change analysis, technical principles, discussion highlights, and risk assessment. | ⭐️⭐️⭐️⭐️⭐️ |
| vllm-rocm-pr-review | Code Review | Review AMD/ROCm-related PRs from vllm (aiter/mori/ROCm kernels) and generate a concise Chinese review report with severity-sorted, type-categorized findings. | ⭐️⭐️⭐️⭐️⭐️ |
| vllm-multimodal-open-issue-analyzer | Issue Analysis | Fetch and organize multimodal-related open issues from vllm. | ⭐️ |
| vllm-technical-blog-writer | Technical Writing | Write or complete Chinese vLLM technical blog posts in the author's established style, with concise architecture, flow, and sequence diagrams. | ⭐️⭐️⭐️⭐️ |
📚 vllm-dev-task-discovery
Prompt:
我想了解 vLLM 的多模态模块中最近还有哪些可以给社区开发者贡献的事情。（Optional：请帮我分析并整理出当前 top 20 的 tasks）
/vllm-dev-task-discovery
Output: vllm_dev_task_multimodal_20260611.
📚 vllm-feature-tutorial
Prompt:
我想了解 vLLM 中的 EPD（Encode-Prefill-Decode）特性，请帮我总结一份技术文档。
/vllm-feature-tutorial
Output: disaggregated_encoder_epd.
📚 vllm-model-tutorial
Prompt:
我想了解 vLLM 中的 DeepSeek-OCR 模型，尤其是其 ViT 部分对多模态输入的处理流程，请帮我总结一份技术文档。
/vllm-model-tutorial
Output: deepseek_ocr.
📚 vllm-feature-design
Prompt:
需求背景：
vLLM 的 EPD 特性目前只支持 ExampleECConnector，我们希望集成 MooncakeECConnector，并通过 Mooncake 来统一管理多种传输后端，比如：TCP/RDMA/SHM。
相关链接：https://github.com/vllm-project/vllm/pull/33714#issuecomment-3882716972
设计目标：
请你参考 vLLM 中 MooncakeConnector 在 PD 分离场景（传输 KV Cache）中的注册和使用方式，实现 MooncakeECConnector，用于支持 EPD 特性下的 Encoder Cache 传输。
目前，你只需要实现 TCP 传输方式，但需要考虑将来可能支持更多的传输后端，比如：RDMA、SHM 等（保证接口的可扩展性），比如：你可以实现 MooncakeECConnector 基类，TCPMooncakeECConnector 继承该基类，未来需要实现其它传输后端时，可以直接添加 RDMAMooncakeECConnector、SHMMooncakeECConnector 等子类。
用户通过 vLLM 启动配置来指定使用哪种 Mooncake 传输后端，不支持启动后切换传输方式，默认使用 TCP 传输。
You also need to rethink how to manage the multi-modal embeddings internally inside vLLM (You might need to introduce "embed blocks" based on hash + position).
注意事项：
你不需要写测试用例，只写核心代码实现，并生成一份 markdown 格式的设计文档，放到当前项目的根目录下。
参考资料：
https://kvcache-ai.github.io/Mooncake/getting_started/supported-protocols.html
https://kvcache-ai.github.io/Mooncake/python-api-reference/transfer-engine.html
/vllm-feature-design
Output: Support Mooncake Based ECConnector for EPD (with manual adjustment and optimization).
📚 vllm-rfc-generator
Prompt:
RFC 标题：
Support ViT Full CUDA Graph
主要内容：
Add full CUDA graph for the ViT to reduce kernel launch overheads.
需要支持图像、视频推理，需要支持 Qwen3-VL/Qwen3.5/GLM-V/Kimi K2.5 等主流 VLM 模型。
相关 PR：
https://github.com/vllm-project/vllm/pull/35963
https://github.com/vllm-project/vllm/pull/37914
https://github.com/vllm-project/vllm/pull/38040
https://github.com/vllm-project/vllm/pull/38061
/vllm-rfc-generator
Output: Support ViT Full CUDA Graph (with manual adjustment and optimization).
📚 vllm-pr-desc-generator
Prompt:
https://github.com/vllm-project/vllm/pull/38061
/vllm-pr-desc-generator
Output: pr-38061-desc (related PR: #38061).
📚 vllm-test-generator
Prompt:
请根据 https://github.com/vllm-project/vllm/pull/38061/changes 中的改动，在 tests/v1/cudagraph/test_encoder_cudagraph.py 中生成对应的测试用例。
要求尽可能复用 SimpleMockViTModel 的代码，减少冗余。
/vllm-test-generator
Output: test_encoder_cudagraph.
📚 vllm-benchmark-result-analysis
Prompt:
Before this PR:
============ Serving Benchmark Result ============
Successful requests: 500
Failed requests: 0
Request rate configured (RPS): 10.00
Benchmark duration (s): 78.58
Total input tokens: 33418
Total generated tokens: 61431
Request throughput (req/s): 6.36
Output token throughput (tok/s): 781.78
Peak output token throughput (tok/s): 2475.00
Peak concurrent requests: 383.00
Total token throughput (tok/s): 1207.07
---------------Time to First Token----------------
Mean TTFT (ms): 7116.24
Median TTFT (ms): 4295.84
P99 TTFT (ms): 18370.87
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 245.78
Median TPOT (ms): 264.03
P99 TPOT (ms): 334.38
---------------Inter-token Latency----------------
Mean ITL (ms): 246.99
Median ITL (ms): 117.71
P99 ITL (ms): 1327.55
==================================================
After this PR:
============ Serving Benchmark Result ============
Successful requests: 500
Failed requests: 0
Request rate configured (RPS): 10.00
Benchmark duration (s): 77.44
Total input tokens: 33418
Total generated tokens: 61522
Request throughput (req/s): 6.46
Output token throughput (tok/s): 794.40
Peak output token throughput (tok/s): 2691.00
Peak concurrent requests: 369.00
Total token throughput (tok/s): 1225.91
---------------Time to First Token----------------
Mean TTFT (ms): 6888.64
Median TTFT (ms): 4128.82
P99 TTFT (ms): 17487.94
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 240.14
Median TPOT (ms): 259.18
P99 TPOT (ms): 313.15
---------------Inter-token Latency----------------
Mean ITL (ms): 241.84
Median ITL (ms): 121.08
P99 ITL (ms): 1470.33
==================================================
/vllm-benchmark-result-analysis
Output: benchmark_comparison_20260323_150830 (related PR: #7104).
📚 vllm-vs-atom-decode-trace-comparison
Prompt:
帮我对比分析 vLLM 和 ATOM 在 MiniMax-M3 上的 decode trace 差异。
模型：MiniMaxAI/MiniMax-M3
vllm trace：/path/to/vllm-minimax-m3-tp4-8k1k-conc8.json
atom trace：/path/to/atom-minimax-m3-tp4-8k1k-conc8.json
报告语言：中文
/vllm-vs-atom-decode-trace-comparison
Output: minimax-m3-trace-comparison_zh_20260905_162716.
📚 vllm-pr-summary
Prompt:
https://github.com/vllm-project/vllm/pull/35963
/vllm-pr-summary
Output: pr-35963-summary.
📚 vllm-rocm-pr-review
Prompt:
https://github.com/vllm-project/vllm/pull/29304
/vllm-rocm-pr-review
Output: pr-29304-review.
📚 vllm-multimodal-open-issue-analyzer
Prompt:
/vllm-multimodal-open-issue-analyzer
Output: vllm_multimodal_issues_20260425_030830.
📚 vllm-technical-blog-writer
Scenario 1 — write a complete article from a topic:
请以“vLLM 中的 Automatic Prefix Caching”为主题，分析最新代码和相关资料，并参考我过往的写作风格生成一篇完整的中文技术博客。请按需补充简洁的架构图和流程图。
/vllm-technical-blog-writer
Scenario 2 — complete an existing Markdown draft:
请基于我提供的“vLLM多模态推理｜ViT_Full_CUDA_Graph\index.md”文章模板和参考资料，结合 vllm-project/vllm 最新 main 分支的代码，按 [Agent TODO] 的要求补充剩余内容。新增图片请放在文章同级的 ./images 目录中。
/vllm-technical-blog-writer
Output: New articles are saved under ./skills/vllm-technical-blog-writer/outputs/<article-slug>/, with referenced images stored in the article's images/ directory. Existing drafts are updated in place.
Coming soon...
通过符号链接将该项目的 skills 目录链接到 ~/.claude/skills/ 下：
ln -sfn /path/to/skill-directory ~/.claude/skills/skill-name
# 示例（新增一个叫 vllm-new-skill 的 skill）：
ln -sfn /Users/shanshan-shen/Documents/GitHub/vllm-dev-skills/skills/vllm-new-skill ~/.claude/skills/vllm-new-skill参数说明：
-s：创建符号链接；-f：如果目标已存在则覆盖（方便更新）；-n：如果目标是已有目录的符号链接，覆盖它而不是跟进到目录内部。
批量链接该目录下所有新增 skill 的命令：
for dir in /Users/shanshan-shen/Documents/GitHub/vllm-dev-skills/skills/*/; do
name=$(basename "$dir")
ln -sfn "$dir" ~/.claude/skills/"$name"
doneTip
@misc{vllm-dev-skills@2026,
title = {vllm-dev-skills},
url = {https://github.com/shen-shanshan/vllm-dev-skills},
note = {Open-source software available at https://github.com/shen-shanshan/vllm-dev-skills},
author = {shen-shanshan},
year = {2026}
}Apache-2.0 License, find more details here.