Skip to content
A/Agent Skills
← Back to catalog

Selection guide / Machine Learning

Know the job.
Choose the tool.

Use one lead skill for the main task. Add a specialist only when it owns a separate phase.

Machine Learning and Data Skills

Every skill in this guide is Claude Code only. For the research workflow around these — experiment design, statistics, and scientific write-up — see research.

Model and training engineering

  • $pytorch-lightning: organize PyTorch code into LightningModules and Trainers, with multi-GPU/TPU, mixed precision, and callbacks. The lead for structuring a training loop; the tools below diagnose one that already runs.
  • $optimize-for-gpu: move scientific Python onto NVIDIA hardware and verify the result is both correct and faster. Pick it for CUDA/GPU acceleration; not for CPU-bound work, where $polars or $dask fit better.
  • $profile-model: read PyTorch profiler traces (.json/.json.gz) to find slow kernels and idle time. Start here when the run is slow but not out of memory.
  • $analyze-memory-snapshot: read CUDA memory snapshots to find OOMs, fragmentation, leaks, and retained tensors. The memory counterpart to $profile-model — reach for it on an out-of-memory failure, not a latency one.
  • $debug-graph-breaks: diagnose and remove torch.compile graph breaks, optionally reaching fullgraph. Narrow to compilation; it does not profile runtime.
  • $abi-stable: migrate C++/CUDA PyTorch extension code to the stable ABI. A one-purpose migration tool, unrelated to training itself.
  • $huggingface-vision-trainer: train or fine-tune vision models (object detection, classification, SAM/SAM2 segmentation) on Hugging Face Jobs cloud GPUs, covering COCO prep, augmentation, evaluation, and Hub persistence. Pick it for vision on HF's managed GPUs; $pytorch-lightning for a training loop you own.
  • $huggingface-trackio: log, alert on, and retrieve ML training metrics with Trackio, with a real-time dashboard. It instruments a training run; $monitor-experiment in research checks an experiment's progress from outside.
  • $get-available-resources: detect host CPU, memory, disk, scheduler, container, and accelerator limits for resource-aware planning. It reports capacity; it does not run or tune anything.

Data and analysis libraries

  • $polars: expression-based DataFrame work with lazy query optimization; the default for fast in-memory ETL and pandas migration.
  • $dask: scale existing pandas/NumPy code past memory or across a cluster. Choose it over $polars only when the data genuinely exceeds one machine — otherwise $polars is faster and simpler.
  • $scikit-learn: classical ML — classification, regression, clustering, pipelines. The lead for non-deep-learning models; use $pytorch-lightning when the model is a neural network.
  • $statsmodels: named statistical models (OLS, GLM, mixed, ARIMA) with full diagnostics. This is the library; $statistical-analysis in research is the guided workflow that decides which test to run.
  • $shap: attribute and audit model predictions with SHAP explainers. Model interpretability, not model fitting.
  • $exploratory-data-analysis: bounded local profiling of supported scientific files (CSV/TSV/JSON, optional NumPy/HDF5). A first-pass survey; hand its findings to a library skill or to $analyze-results for the experiment loop.

On-device AI (Apple)

Ships in the apple-skills plugin. These target Apple's on-device model frameworks, not server-side training — reach for the sections above to train or fine-tune a model, and these to ship inference inside an Apple app.

  • $core-ml: Core ML, Create ML, Vision, and Natural Language for on-device image classification, text analysis, and model integration.
  • $apple-intelligence: Apple Intelligence features — Foundation Models, Visual Intelligence, App Intents, and intelligent assistants.

Visualization libraries

  • $matplotlib: fine-grained control over every plot element and novel plot types. The low-level lead; verbose for routine statistical plots.
  • $seaborn: quick, attractive statistical plots with pandas integration for distributions and categorical comparisons. Prefer it for standard exploratory charts; drop to $matplotlib when you need control it cannot give. For publication figures held to a truthfulness and accessibility bar, use $scientific-visualization in research.

Geospatial

  • $geopandas: GeoSeries/GeoDataFrame workflows, spatial operations, and vector I/O, with local audit tools.
  • $geomaster: broad geospatial science — remote sensing, GIS, spatial analysis, and ML for earth observation. Pick $geopandas for a concrete vector-data task; $geomaster for the wider earth-observation problem.

Project-specific (SiamYOLO)

These are bound to the SiamYOLO / DisasterAdaptiveNet repository and answer questions about that model. They do not generalize to other codebases.

  • $model-architecture-analyzer: trace the Siamese YOLO architecture end to end — 6-channel input, dual weight-shared backbones, bitemporal fusion, neck, heads.
  • $tensor-shape-tracer: trace tensor shapes through that model — per-branch slicing, backbone strides, fusion channel math.
  • $loss-function-auditor: prove which loss the training run actually computes (box/cls/dfl, mask, ordinal terms).
  • $baseline-comparator: compare SiamYOLO against stock Ultralytics and the DisasterAdaptiveNet reference; keep the split and repo attached to every score.
  • $paper-code-mapper: map thesis methodology to the actual code and back (FiLM conditioning, ComboLoss, xBD).