Claude Skills by mxslr

Use when an existing model's results are disappointing and the user wants higher accuracy - 'accuracy is still too low', 'improve/boost the model', 'why is it bad', 'make it better than the paper', 'combine methods to beat SOTA'. Diagnoses the real bottleneck first (error analysis, learning curves, leakage, saliency), then researches the specific gap, synthesizes a principled combination of techniques from the literature, and ranks next levers by expected impact - with honest ceilings, not ma...

Use BEFORE training any model, to build correct train/val/test splits and hunt data leakage - the #1 cause of fake-high accuracy. Covers group/patient/subject splits, temporal splits, official-benchmark splits, label correctness, class balance, and preprocessing parity. Triggers on 'split the data', 'train/test split', 'cross validation', 'why is accuracy suspiciously high', 'data leakage', 'per-patient'.

Use as the FIRST step of any ML task, before choosing a model, to inspect and understand the actual dataset. Works for a LOCAL dataset (Claude reads the files directly) and for a KAGGLE dataset (Claude cannot read /kaggle/input from your machine, so it emits a small profiling cell you run on Kaggle and paste back, or downloads via the Kaggle API). Profiles structure, size, modality, schema, class balance, the entity for group splitting, and any official split. Triggers on 'my dataset', 'data ...

Use when turning a trained model into something usable (an app, an inference function, a demo) or adding explainability. Selects the RIGHT interpretability method per problem: Grad-CAM for CNN image classification, predicted boxes for detection, mask overlays for segmentation, embedding neighbors for face recognition, SHAP for tabular, attributions or attention for text, feature and temporal importance for time-series. Also covers preprocessing parity between training and inference, calibrate...

Use for 3D and point cloud tasks: point cloud classification, 3D semantic or instance segmentation, 3D object detection from LiDAR, and depth or mesh analysis, including autonomous-driving and indoor-scan settings. Picks the right 3D backbone, representation, and scene-level split and metric. Triggers on 'point cloud', '3D', 'LiDAR', 'mesh', 'depth', '3D detection', '3D segmentation', 'ScanNet', 'KITTI', 'ModelNet', 'autonomous driving perception'.

Use for anomaly, outlier, and novelty detection across data types: industrial or visual defect detection, time-series or sensor anomalies and predictive maintenance, tabular fraud or intrusion detection, and out-of-distribution detection. Usually one-class or unsupervised. Picks the method by data type, sets the threshold honestly, and uses imbalance-aware metrics. Triggers on 'anomaly detection', 'outlier', 'defect detection', 'novelty', 'fraud', 'intrusion', 'out of distribution', 'predicti...

Use for audio and speech tasks: speech recognition (speech to text, ASR, transcription), audio or sound classification and tagging, keyword spotting, speaker identification or verification, sound event detection, and music or audio analysis. Picks the best model per sub-task and the right features, splits, and metrics. Triggers on 'speech to text', 'transcribe', 'ASR', 'audio classification', 'sound', 'keyword spotting', 'speaker recognition', 'voice', 'audio dataset'.

Use for computer-vision tasks on natural/general images: image classification, object detection (vehicles, faces, people, products), face recognition/verification, semantic/instance segmentation, keypoints/pose, OCR. Picks the best architecture per sub-task and the right transfer-learning + augmentation recipe. Triggers on 'detect objects/faces/vehicles', 'classify images', 'segment', 'bounding boxes', 'recognize faces', 'image dataset'.

Use for generative modeling of images and audio: text-to-image, image editing or image-to-image, unconditional image generation, inpainting, super-resolution, style transfer, and audio or music generation, including fine-tuning or personalizing a base model. Picks diffusion vs GAN, the right adaptation method, and generation-specific evaluation. For text generation use domain-nlp-llm instead. Triggers on 'generate images', 'text to image', 'diffusion', 'stable diffusion', 'GAN', 'image editin...

Use for graph machine learning: node classification, link prediction, graph classification or regression, recommendation on graphs, molecule or network analysis, community detection, and graph anomaly detection with graph neural networks. Picks the GNN by task, enforces transductive vs inductive splits (a common leakage source), and uses graph metrics. Triggers on 'graph neural network', 'GNN', 'node classification', 'link prediction', 'knowledge graph', 'molecule', 'social network', 'GCN', '...

Use for medical-image AI: tumor/cancer detection & classification, brain tumor MRI, mammography (benign vs malignant), chest X-ray, CT, histopathology, retinal/fundus, ultrasound, dermoscopy. Encodes hard-won rigor: patient-level splits (no leakage), medical preprocessing (CLAHE, ROI/organ cropping, artifact/pectoral removal), clinical operating points (high sensitivity), probability calibration, and skepticism toward inflated ROI-crop/leaky benchmark numbers. Triggers on 'detect cancer/tumor...

Use for vision-language and multimodal tasks: image-text retrieval, zero-shot image classification, visual question answering (VQA), image captioning, document or chart understanding, and visual grounding. Picks a contrastive dual-encoder or an instruction-tuned VLM, and uses task-appropriate evaluation. Triggers on 'vision language', 'image and text', 'CLIP', 'VQA', 'visual question answering', 'image captioning', 'multimodal', 'describe an image', 'document understanding', 'zero-shot image'.

Use for text / NLP / LLM tasks: text classification, sentiment, NER/token tagging, QA, summarization, translation, semantic search / embeddings, retrieval-augmented generation (RAG), chatbots/agents, and fine-tuning language models. Picks prompt vs fine-tune vs RAG, the right model, and rigorous text evaluation. Triggers on 'classify text', 'sentiment', 'NER', 'summarize', 'chatbot', 'RAG', 'embeddings', 'fine-tune LLM', 'question answering'.

Use for recommendation and ranking systems: product or content recommendation, collaborative filtering, candidate retrieval, learning to rank, next-item or sequential recommendation, and cold start. Picks retrieve-then-rank architecture, the right model, honest temporal splits, and ranking metrics. Triggers on 'recommendation', 'recommender', 'collaborative filtering', 'ranking', 'personalization', 'next item', 'user-item', 'what to show users'.

Use for reinforcement learning and sequential decision making: game or control agents, robotics, continuous or discrete control, offline RL from logged data, and contextual bandits. Picks the algorithm by setting, defines reward and state carefully, and evaluates with proper multi-seed protocol. For aligning an LLM (RLHF) use domain-nlp-llm. Triggers on 'reinforcement learning', 'RL agent', 'policy', 'reward', 'control', 'robotics', 'game playing', 'PPO', 'DQN', 'bandit'.

Use for structured/tabular data prediction: churn, fraud, credit/risk scoring, price/demand regression, conversion, recommendation features, any CSV/database ML. Picks gradient boosting vs deep tabular, robust feature engineering, correct CV, calibration, and imbalance handling. Triggers on 'predict churn/fraud/price/risk', 'classify this CSV/table', 'tabular', 'XGBoost/LightGBM', 'feature engineering'.

Use for time-series tasks: forecasting (demand, sales, energy, finance), time-series classification, and anomaly detection over time (sensors, logs, monitoring). Picks classical vs ML vs deep forecasters, enforces temporal splits (no future leakage), and uses proper time-series metrics/backtesting. Triggers on 'forecast', 'predict next/future', 'time series', 'demand/sales/energy prediction', 'anomaly detection over time', 'seasonality'.

Use for video understanding: action or activity recognition, video classification, temporal action detection or localization in untrimmed video, gesture recognition, and general video analysis. Picks the right video model and the correct video-level split and metric. Triggers on 'video classification', 'action recognition', 'activity recognition', 'temporal action detection', 'video understanding', 'recognize actions', 'gesture', 'video dataset'.

Use when starting an ML task or choosing a method, to research papers/benchmarks and find the best approach and the REALISTIC accuracy ceiling. Finds SOTA, critically appraises reported numbers, and flags claims inflated by data leakage or ROI-cropping. Triggers on 'what's the best model/method for', 'state of the art', 'research papers', 'benchmark', 'how accurate can this get'.

Use at the START of ANY machine-learning / deep-learning / AI modeling task - building, training, fine-tuning, or choosing a model for image classification, object/face/vehicle detection, segmentation, medical imaging (tumor/cancer/MRI/X-ray/mammogram), text/NLP/LLM, tabular prediction (churn, price, risk), or time-series forecasting. Orchestrates the professional research-engineer workflow: research papers first, pick the best method, train rigorously, evaluate honestly (no data leakage, no ...

Use when producing a runnable deliverable for an ML task, to decide the target environment and author the notebook or script. Detects whether the user wants a Kaggle notebook, a Google Colab notebook, or a local GPU run, and adapts paths, data loading, and compute settings accordingly. Also enforces the writing style for generated notebooks: short and simple explanations, professional research tone, no em-dash, no emoji. Triggers on 'make a notebook', 'kaggle', 'google colab', 'run on my loca...

Use when evaluating or reporting model performance and choosing metrics, thresholds, and plots that FIT the problem type. Picks the right metrics per task (binary uses ROC-AUC and PR-AUC; multiclass uses macro F1 and a confusion matrix; detection uses mAP; segmentation uses Dice and IoU; regression uses MAE, RMSE, R2; ranking uses nDCG) and does not force AUC or ROC where they do not apply. Also covers calibration, thresholds chosen on validation, confidence intervals, and honest reporting. T...

Use when training or fine-tuning a deep model and tuning it for best accuracy. Covers transfer learning, two-phase fine-tuning, discriminative/layer-wise LR, LR schedules + warmup, regularization tuned to over/under-fit (dropout, weight decay, label smoothing, MixUp), mixed precision, gradient checkpointing, EMA, checkpoint/resume, and fitting a GPU/time budget. Triggers on 'fine-tune', 'train the model', 'best hyperparameters', 'overfitting', 'learning rate', 'out of memory', 'training is to...