MLOps
模型部署、评估与运维
model-interpretability
levy-n
Model interpretability, explainability, and debugging tools. Covers SHAP (TreeExplainer, DeepExplainer, KernelExplainer), feature importance analysis, LIME, attention visualization, Grad-CAM for CNNs, confusion matrix analysis, error analysis patterns, and model fairness auditing. Use when user asks about 'SHAP', 'feature importance', 'explainability', 'interpretability', 'why did the model predict', 'Grad-CAM', 'LIME', 'attention weights', 'confusion matrix', 'error analysis', 'model debugging', 'fairness', 'bias detection', or 'what did the model learn'.
sequence-models
levy-n
Implements sequence models for time series and text. Covers RNN fundamentals, LSTM/GRU architectures, time series forecasting, text generation with language models, and sequence classification. Use when working with sequential data, predicting time series, text generation, or when user mentions 'RNN', 'LSTM', 'GRU', 'vanishing gradient', 'hidden state', 'time series', 'sequence-to-sequence', 'text generation', 'next word prediction', or 'recurrent neural network'.
fine-tuning-peft
levy-n
Expert guide for LLM fine-tuning and parameter-efficient training methods. Covers LoRA, QLoRA, PEFT library, adapter tuning, instruction tuning, quantization (GPTQ, AWQ, GGUF, bitsandbytes), dataset preparation for fine-tuning, Hugging Face Trainer/TRL, RLHF/DPO/ORPO alignment, and model merging. Use when user asks about 'fine-tuning', 'LoRA', 'QLoRA', 'PEFT', 'adapter', 'quantization', 'bitsandbytes', '4-bit', '8-bit', 'instruction tuning', 'RLHF', 'DPO', 'model merging', 'Unsloth', 'Axolotl', 'training custom models', 'TRL', or 'SFT'.
generative-models
levy-n
Generative AI models: GANs, VAEs, Diffusion Models, and image generation. Covers GAN architecture (Generator/Discriminator), DCGAN, Wasserstein GAN, Variational Autoencoders, latent space interpolation, Diffusion models (DDPM), Stable Diffusion, conditional generation, and text-to-image. Use when user asks about 'GAN', 'generative adversarial', 'VAE', 'variational autoencoder', 'diffusion model', 'image generation', 'Stable Diffusion', 'DCGAN', 'Wasserstein', 'WGAN', 'latent space', 'generate images', 'text-to-image', 'DDPM', 'denoising diffusion', 'style transfer', or 'deepfake'.
tsfm-forecast
dtsong
"Use this skill when generating time-series forecasting pipelines using foundation models. Covers TimesFM, Chronos, MOIRAI, and Lag-Llama model selection, DuckDB-based preprocessing code, Python inference generation, backtesting harnesses, multi-model comparison, and client forecast deliverables. Common phrases: \"time-series forecast\", \"demand forecasting\", \"TimesFM\", \"Chronos\", \"predict future values\", \"zero-shot forecast\". Do NOT use for ML model training or fine-tuning (use python-data-engineering), real-time/streaming forecasts (use event-streaming), or pipeline scheduling (use data-pipelines)."
transformers
eyadsibai
Use when "HuggingFace Transformers", "pre-trained models", "pipeline API", or asking about "text generation", "text classification", "question answering", "NER", "fine-tuning transformers", "AutoModel", "Trainer API"
SUITE_NAME
dtsong
TRIGGER_DESCRIPTION. Use when USER_CONTEXT. Routes to specialists for CAPABILITIES.
llm-as-a-judge
maragudk
Build, validate, and deploy LLM-as-Judge evaluators for automated quality assessment of LLM pipeline outputs. Use this skill whenever the user wants to: create an automated evaluator for subjective or nuanced failure modes, write a judge prompt for Pass/Fail assessment, split labeled data for judge development, measure judge alignment (TPR/TNR), estimate true success rates with bias correction, or set up CI evaluation pipelines. Also trigger when the user mentions "judge prompt", "automated eval", "LLM evaluator", "grading prompt", "alignment metrics", "true positive rate", or wants to move from manual trace review to automated evaluation. This skill covers the full lifecycle: prompt design → data splitting → iterative refinement → success rate estimation.
data-science
eyadsibai
Use when "statistical modeling", "A/B testing", "experiment design", "causal inference", "predictive modeling", or asking about "hypothesis testing", "feature engineering", "data analysis", "pandas", "scikit-learn"
business-model-preferences
phrazzld
Pricing philosophy and business model constraints. Auto-invoke when: evaluating pricing, checkout flows, subscription logic, tier structures.
transformers-llm
levy-n
Implements Transformer models and LLM workflows. Covers attention mechanism, BERT fine-tuning, HuggingFace Transformers library (Tokenizer, Trainer, Pipeline), and LLM ecosystem (GPT, Claude, Gemini, Ollama). Use when fine-tuning language models, using HuggingFace, calling LLM APIs, or when user mentions 'transformer', 'attention', 'BERT', 'HuggingFace', 'tokenizer', 'fine-tuning', 'LLM', 'GPT', 'Claude', 'Gemini', 'prompt engineering', 'zero-shot', or 'few-shot learning'.
debug:scikit-learn
SnakeO
Debug Scikit-learn issues systematically. Use when encountering model errors like NotFittedError, shape mismatches between train and test data, NaN/infinity value errors, pipeline configuration issues, convergence warnings from optimizers, cross-validation failures due to class imbalance, data leakage causing suspiciously high scores, or preprocessing errors with ColumnTransformer and feature alignment.
speckit-tasks-zh
forztf
基于speckit工作流的任务生成技能，用于根据可用设计文档生成可操作的、依赖有序的tasks.md。当需要基于spec.md、plan.md、data-model.md、contracts/等设计文档为功能开发生成详细任务列表时使用此技能。触发词包括"speckit tasks"、"生成任务"、"任务规划"、"功能任务分解"、"创建tasks.md"等。
debug:tensorflow
SnakeO
Debug TensorFlow and Keras issues systematically. This skill helps diagnose and resolve machine learning problems including tensor shape mismatches, GPU/CUDA detection failures, out-of-memory errors, NaN/Inf values in loss functions, vanishing/exploding gradients, SavedModel loading errors, and data pipeline bottlenecks. Provides tf.debugging assertions, TensorBoard profiling, eager execution debugging, and version compatibility guidance.
refactor:scikit-learn
SnakeO
Refactor Scikit-learn and machine learning code to improve maintainability, reproducibility, and adherence to best practices. This skill transforms working ML code into production-ready pipelines that prevent data leakage and ensure reproducible results. It addresses preprocessing outside pipelines, missing random_state parameters, improper cross-validation, and custom transformers not following sklearn API conventions. Implements proper Pipeline and ColumnTransformer patterns, systematic hyperparameter tuning, and appropriate evaluation metrics.
debug:pytorch
SnakeO
Debug PyTorch issues systematically. Use when encountering tensor errors, CUDA out of memory errors, gradient problems like NaN loss or exploding gradients, shape mismatches between layers, device conflicts between CPU and GPU, autograd graph issues, DataLoader problems, dtype mismatches, or training instabilities in deep learning workflows.
refactor:pytorch
SnakeO
Refactor PyTorch code to improve maintainability, readability, and adherence to best practices. Identifies and fixes DRY violations, long functions, deep nesting, SRP violations, and opportunities for modular components. Applies PyTorch 2.x patterns including torch.compile optimization, Automatic Mixed Precision (AMP), optimized DataLoader configuration, modular nn.Module design, gradient checkpointing, CUDA memory management, PyTorch Lightning integration, custom Dataset classes, model factory patterns, weight initialization, and reproducibility patterns.
skills
atxinsky
Execute plan in batches with review checkpoints
transformers-js
nico-martin
Use Transformers.js to run state-of-the-art machine learning models directly in JavaScript/TypeScript. Supports NLP (text classification, translation, summarization), computer vision (image classification, object detection), audio (speech recognition, audio classification), and multimodal tasks. Works in Node.js and browsers (with WebGPU/WASM) using pre-trained models from Hugging Face Hub.
nuxt-enums
leeovery
TypeScript enum pattern with Castable interface for model integration. Use when creating enums with behavior methods (colors, labels), defining fixed value sets, or integrating enums with the model casting system.
azure-ai-formrecognizer-java
Tryboy869
"Build document analysis applications with Azure Document Intelligence (Form Recognizer) SDK for Java. Use when extracting text, tables, key-value pairs from documents, receipts, invoices, or building custom document models."
jackyshen-design-workshop-outline
mebusw
Use when user asks to "generate workshop outline", "create training agenda", "design course structure", "build workshop schedule", or requests help planning training sessions. Applies MECE structure, TfBR design (4Cs), and VAK-inclusive learning.
nuxt-layers
leeovery
Working with Nuxt layers (base, nuxt-ui, x-ui) that provide shared functionality. Use when understanding layer architecture, importing from layers, extending layer functionality, or creating new layers.
capa-officer
QuestNova502
Senior CAPA Officer specialist for managing Corrective and Preventive Actions within Quality Management Systems. Provides CAPA process management, root cause analysis, effectiveness verification, and continuous improvement coordination. Use for CAPA investigations, corrective action planning, preventive action implementation, and CAPA system optimization.