ml-skills

ML Skills

A Claude Code plugin that ships one skill — ml-review — and a curated reference library of ~80 ML topics. The skill routes broad ML/DL questions, reviews ML approaches, analyzes pipelines, and suggests solutions; the references hold the deep treatment of each topic and are read on demand by the skill.

Author: hung-phan

Install

Register as a plugin marketplace, then install:

/plugin marketplace add hung-phan/ml-skills
/plugin install ml-skills@ml-skills

Local install (for development)

claude plugin marketplace add ./
claude plugin install ml-skills@ml-skills

How to Use

There is one skill: /ml-review. Reach for it whenever you have an ML/DL question, a plan to critique, or a system that's misbehaving.

/ml-review I need to fine-tune Llama-3 on a single GPU with limited VRAM
/ml-review my XGBoost model passes CV but fails in production — what do I check?
/ml-review review my plan: train a transformer on 50K rows of tabular data
/ml-review p95 TTFT is 800ms on vLLM with Llama-3-8B — what can I tune?
/ml-review which architecture should I use for ultra-long-sequence forecasting?

/ml-review reads your question and does one of three things:

1. Routes — points you to the right reference doc (e.g. references/ml-architectures/attention/).
2. Reviews — critiques an ML plan against the reference library, surfacing the top 2-3 risks.
3. Analyzes / suggests — diagnoses a system or proposes an end-to-end approach with citations.

It does this by reading reference files on demand — the references are not preloaded into context.

Repository Layout

.claude-plugin/plugin.json           # plugin manifest (declares the single ml-review skill)
skills/
  ml-review/
    SKILL.md                         # the only auto-loaded skill — routing tables + review/analyze/suggest procedures
    references/                      # ~80 reference docs, read on demand by ml-review
      ml-architectures/              # Agents, Attention, Audio, CNN, Diffusion, LLM, Mamba, MoE, RAG, RL, Vision, World Models, …
      ml-libraries/                  # PyTorch, HuggingFace, sklearn, XGBoost, pandas, polars, numpy, Ray, vLLM, SGLang, …
      ml-training/                   # training-workflow, evaluation, llm-evaluation, prompt-engineering, inference-optimization, model-merging, online-experimentation, online-learning, data-parallel, Unsloth, GRPO, …
      data-prep/                     # eda, feature-engineering, time-series-features, dataset-curation, data-validation
      gpu-lang/                      # triton, tilelang
docs/
  acquire-ml-skill.md                # maintainer guide: add or update a single reference
  refine-ml-skill.md                 # maintainer guide: deep-research + propagate updates across multiple references
README.md

The references/ folder follows Claude's skill convention — files there are not auto-loaded into context but the skill reads them with the Read tool when needed. This keeps the per-session token footprint small.

What lives in references/

Folder	Topics
ml-architectures/	Agents, AI App Architecture, ANN, Attention, Audio, Autoencoder, Boltzmann, CNN, Diffusion, Embeddings, GAN, GNN, LLM, Mamba, MoE, Neural Combinatorial Optimization, Quantization, RAG, Regression/Classification, Reinforcement Learning, RNN, Sampling Strategies, SOM, Transformer, Vision, World Models
ml-libraries/	PyTorch, HuggingFace, scikit-learn, XGBoost, pandas, polars, numpy, Ray, NeMo, DSPy, LiteLLM, vLLM, SGLang, Triton Inference Server, keras, seaborn, plotly
ml-training/	feature-selection, training-workflow, evaluation (classical) + llm-evaluation (FM-specific), prompt-engineering, inference-optimization, model-merging, gradient-free-optimization, experiment-tracking, hf-jobs-workflow, online-experimentation (A/B, CUPED, bandits), online-learning (drift, incremental updates), data-parallel (DDP/FSDP), Unsloth SFT, Unsloth advanced (GRPO/DPO), Ray distributed SFT, distributed GRPO
data-prep/	EDA, feature engineering, time-series-features (lags, windows, point-process, purged CV), dataset-curation (FM data: SFT, preference, CoT, synthesis, dedup), data validation
gpu-lang/	Triton, TileLang

How the Skill Picks References

ml-review uses three lookup paths, in this order of preference:

1. Workflow-stage table — when you know what step you're on (e.g. "I'm picking metrics" → references/ml-training/evaluation/).
2. Problem-type tables — when you know what kind of problem you have (e.g. "tabular forecasting" → references/data-prep/time-series-features/).
3. grep — when you know a keyword (paper name, API, library) but not which reference owns it.