AI Infrastructure Agent Skills

⚠️ WARNING This project is under active development and heavily generated by LLMs without strict proofreading. Use with caution and verify all code before production use.

A collection of specialized agent skills for AI infrastructure engineers, covering both technical development (GPU kernels, distributed training, inference optimization) and soft skills (flowchart creation, presentation design).

Overview

This repository provides expert-level skills tailored for AI infrastructure engineering. Each skill packages domain knowledge, code examples, and best practices to transform Claude into a specialized assistant for specific frameworks and workflows—from writing high-performance CUDA kernels to creating professional technical presentations.

Construction Methodology (Unless Otherwise Specified)

1. Knowledge Gathering: Use Gemini DeepResearch to collect comprehensive, up-to-date information on target frameworks
2. Skill Development: Transform research into structured skills using skill-creator in Claude Code
3. Validation: Test skill-generated code examples to ensure correctness
4. Maintenance: Regular updates based on latest official documentation

Available Skills

TileLang Developer

Write high-performance GPU kernels using TileLang for NVIDIA, AMD, and Ascend hardware.

Capabilities:

• Matrix multiplication (GEMM) kernels
• FlashAttention implementations
• DeepSeek MLA operators
• Performance optimization (swizzle layouts, pipelining, warp specialization)
• Cross-platform kernel development

Status: ✅ Complete

Megatron Memory Estimator

Estimate GPU memory usage for Megatron-based MoE and dense models. Built upon megatron_memory_estimator.

Capabilities:

• Estimate memory from HuggingFace configs
• Support for MoE models (DeepSeek-V3, Qwen, etc.)
• Parallelism strategy comparison (TP/PP/EP/CP)
• Memory optimization recommendations

Status: ✅ Complete

SLIME User

Guide for using SLIME (LLM post-training framework for RL Scaling). Built upon THUDM/slime.

Capabilities:

• RL training setup and configuration (GRPO, GSPO, PPO, Reinforce++)
• Multi-turn tool calling and agent workflows
• Custom reward models and generation functions
• Megatron and FSDP backend configuration
• SGLang integration and optimization
• Dynamic sampling and partial rollout
• Multi-node distributed training

Status: ✅ Complete

Prompt to create this skill, with Sonnet 4.5:

Use skill-creator to create a skill called slime-user at this repo. slime is an LLM
post-training framework for RL Scaling. Its repo is https://github.com/THUDM/slime.

Skill creation procedure:

1. Git clone the latest repo
2. Analyze `docs/en`, understand basic structure and write a doc navigation guide for user
getting started or finding docs for advanced usage
3. Gather valuable examples from the docs and `examples` dir, write key ideas and script
path down for quick reference
4. Checkout some important source code, for example `slime/slime/utils/arguments.py` and
`slime/rollout/sglang_rollout.py`, provide its path and functions for a quick find.

TikZ Flowchart

Create professional flowcharts and architecture diagrams using LaTeX TikZ with standardized styles.

Capabilities:

• Professional flowcharts with Google Material-like color palette
• Standardized node types (data, memory, operation, kernel boxes)
• Architecture diagrams and process flows
• Grouping and layout best practices
• Clean orthogonal edges and relative positioning

Example Output: QAT Flowchart | Anthropic Theme

Status: ✅ Complete

Material You Slides

Create presentation slide decks using Material You (Material Design 3) design language.

Capabilities:

• Self-contained HTML slides (1280x720) with M3 color tokens
• Roboto typography with multiple weight support
• Professional slide types (title, section divider, content)
• Component library (cards, flow diagrams, metric cards, code blocks)
• Rounded shapes and generous whitespace
• Surface hierarchy without drop shadows
• Structured layouts (columns, tables, lists, tags/chips)

Example Output: SLIME RL Training Slides

Status: ✅ Complete

Anthropic Theme Flowchart

Create polished standalone HTML/CSS flowcharts with Anthropic-inspired pastel styling, reliable geometry, and deterministic connector routing.

Capabilities:

• TS-first flowchart specs that generate standalone HTML artifacts
• Deterministic node geometry and anchor-based connector routing
• Transparent dashed grouping frames and pastel role-based node styling
• Hollow > arrowheads and orthogonal bridge connectors
• Gallery-ready exports generated from the same geometry source

Example Output: Review Demo PNG

Status: ✅ Complete

HF Architecture TikZ

Generate Sebastian-Raschka-gallery-style TikZ architecture diagrams for any HuggingFace decoder-only LLM, with per-block parameter formulas and concrete numbers.

Capabilities:

• Extract architecture from HuggingFace configs into a structured spec
• Render publication-quality vertical TikZ diagrams via Jinja templates
• Annotate every sub-block with parameter formulas and concrete numbers
• Support MHA, GQA, MLA, Hyper-Connections, sparse attention with learned indexer
• Cover dense and MoE FFNs (incl. hash routing) and MTP heads
• Models: DeepSeek-V4-Flash, Qwen, Llama, Mistral, gpt-oss, etc.

Example Output: DeepSeek-V4-Flash PNG | PDF

Status: ✅ Complete

OpenAI Dotcom Viz

Build figures in OpenAI's blog / research / system-card "dotcom" visual style, emitted as self-contained HTML/SVG with zero dependencies.

Capabilities:

• Monochrome bar charts (darker same-hue stroke, rounded corners, black y-axis with outward ticks, no gridlines, circle legend markers, value labels above bars, angled category labels)
• Flow / process diagrams (rounded boxes, monospace uppercase pills, pink highlights, thin open-chevron connectors, dashed negative branches)
• Real OpenAI Sans typography with zero external dependencies

Status: ✅ Complete

Planned Skills

SGLang Developer

Development skill for SGLang (Structured Generation Language) runtime and optimization.

Planned capabilities:

• SGLang runtime configuration
• Custom sampling strategies
• Performance tuning for LLM inference
• Multi-GPU serving optimization

Status: 🚧 Planned

vLLM Developer

Skill for vLLM engine development and deployment.

Planned capabilities:

• PagedAttention implementation
• Custom scheduler development
• Multi-LoRA serving
• Quantization integration

Status: 🚧 Planned

Usage

Installing Skills

Skills are installed by placing the skill directory in Claude's skills path:

Natural Language: Ask Claude Code directly: "Help me install skills from https://github.com/yzlnew/infra-skills"

Personal (across all projects):

# Clone and copy to personal skills directory
git clone https://github.com/yzlnew/infra-skills.git
mkdir -p ~/.claude/skills
cp -r infra-skills/tilelang-developer ~/.claude/skills/
cp -r infra-skills/megatron-memory-estimator ~/.claude/skills/
cp -r infra-skills/slime-user ~/.claude/skills/
cp -r infra-skills/tikz-flowchart ~/.claude/skills/
cp -r infra-skills/material-you-slides ~/.claude/skills/
cp -r infra-skills/anthropic-theme-flowchart ~/.claude/skills/
cp -r infra-skills/hf-architecture-tikz ~/.claude/skills/

Project-level (for repository collaborators):

# Clone and copy to project's skills directory
cd your-project
git clone https://github.com/yzlnew/infra-skills.git .claude/skills-repo
mkdir -p .claude/skills
cp -r .claude/skills-repo/tilelang-developer .claude/skills/
cp -r .claude/skills-repo/megatron-memory-estimator .claude/skills/
cp -r .claude/skills-repo/slime-user .claude/skills/
cp -r .claude/skills-repo/tikz-flowchart .claude/skills/
cp -r .claude/skills-repo/material-you-slides .claude/skills/
cp -r .claude/skills-repo/anthropic-theme-flowchart .claude/skills/
cp -r .claude/skills-repo/hf-architecture-tikz .claude/skills/

Skills automatically activate when relevant tasks are detected.

Examples

TileLang Kernel Development:

# User request:
"Write a FP16 matrix multiplication kernel optimized for A100"

# Claude loads tilelang-developer skill and generates:
# - Complete TileLang kernel code
# - Performance optimizations (swizzle, pipelining)
# - Testing code
# - Hardware-specific tuning recommendations

Megatron Memory Estimation:

# User request:
"Estimate memory for DeepSeek-V3 with TP=8, PP=4, EP=8"

# Claude loads megatron-memory-estimator skill and provides:
# - Detailed memory breakdown (model, optimizer, activations)
# - Comparison across different parallelism strategies
# - Memory optimization recommendations
# - Hardware configuration suggestions

SLIME RL Training Setup:

# User request:
"Help me set up GRPO training for Qwen3-4B with multi-turn tool calling"

# Claude loads slime-user skill and provides:
# - Environment setup instructions
# - Custom generation function for tool calling
# - Training script configuration
# - Multi-node scaling guidance

TikZ Flowchart Creation:

# User request:
"Create a flowchart showing the FlashAttention-2 algorithm flow"

# Claude loads tikz-flowchart skill and generates:
# - Professional LaTeX TikZ diagram with standardized colors
# - Data nodes (green), operation nodes (blue), memory nodes (orange)
# - Clean layout with orthogonal edges
# - Grouped kernel phases with proper styling

Material You Slides Creation:

# User request:
"Create a presentation deck about our AI infrastructure architecture"

# Claude loads material-you-slides skill and generates:
# - Self-contained HTML file with Material Design 3 styling
# - Title slide with gradient background and branding
# - Section dividers with large translucent numbers
# - Content slides with cards, flow diagrams, and metric displays
# - Responsive 1280x720 slides ready for presentation

Development

Testing Skills

Validate code examples in skills:

# Run all tests from project root
pytest

# Run tests for specific skill
pytest tests/tilelang-developer/

# Run specific test file
pytest tests/tilelang-developer/test_gemm.py

Updating Skills

When frameworks release major updates:

1. Update skill source files (SKILL.md, references/) with latest information
2. Run validation tests to ensure examples are correct
3. Commit and tag new version

Quality Standards

All skills must meet these criteria:

• ✅ Accurate: Code examples must be tested and correct
• ✅ Concise: Follow progressive disclosure (SKILL.md < 500 lines)
• ✅ Complete: Include workflow, API reference, examples, and debugging
• ✅ Current: Based on latest stable framework version
• ✅ Clear: Explicit triggers in description for automatic activation

Contributing

Skill Requests

Open an issue with:

• Framework/tool name
• Use cases and scenarios
• Link to official documentation

Skill Improvements

1. Fork the repository
2. Update skill source files
3. Run validation tests
4. Submit PR with changelog

Roadmap

• TileLang developer skill
• Megatron memory estimator skill
• SLIME user skill
• TikZ flowchart skill
• Material You slides skill
• Anthropic theme flowchart skill
• HF architecture TikZ skill
• SGLang developer skill
• vLLM developer skill
• Automated testing pipeline
• Documentation update monitoring
• Skill versioning system

Resources

• Claude Agent Skills Documentation
• skill-creator Guide
• TileLang Repository
• SLIME Repository

License

Skills are provided as-is for development purposes. Generated code follows the license terms of the underlying frameworks.

---

Note: This is a specialized repository for AI infrastructure developers. Skills contain advanced technical content and assume familiarity with GPU programming, compiler design, and deep learning systems.