pytorch/pytorch's Claude Code Skills for Core Contributors

What is the pytorch/pytorch plugin, actually?

pytorch/pytorch is the source repository for PyTorch itself, the deep learning framework GitHub currently credits with 102,589 stars by Flowy's own count. Buried inside that same repository, alongside the framework's C++ and Python source, its maintainers keep a folder of Claude Code skills built for a different job: not training models with PyTorch, but developing PyTorch's own codebase. Flowy indexes 16 skills from it, and all 16 are written for someone already working inside that source tree.

If your day job is calling PyTorch's ordinary API to build and train neural networks, none of these 16 skills touch that surface. They touch issue triage, pull request review, kernel code, the PyTorch 2 compiler stack, and the project's own documentation and CI dashboards: the parts of the job that belong to the people who maintain PyTorch rather than the people who use it.

How does it help with issues and pull requests?

Five of the sixteen skills cover the contribution workflow end to end. triaging-issues routes a freshly filed GitHub issue to the right oncall team, applies labels, and closes questions that do not need engineering work. distributed-triage picks up from there for anything already routed to the distributed oncall queue, assigning module labels and sub-routing it further. scrub-issue fetches a bug report and tries to reproduce and minimize it, so a maintainer knows a repro actually holds before spending time on it. Once a fix is in hand, fix-issue reproduces the reported bug, root-causes it, and implements a fix in the local working tree. pr-review closes the loop, reviewing a pull request for code quality, test coverage, security, and backward compatibility.

What about kernels, dispatch macros, and the compiler stack?

A second group of skills works much closer to the hardware, on the C++ and CUDA code that makes PyTorch's operators actually run. add-uint-support adds unsigned integer types, such as uint16, uint32, and uint64, to an operator by updating its AT_DISPATCH macros. at-dispatch-v2 converts the older dispatch macro style used across ATen kernels into the newer v2 form. metal-kernel writes Metal and MPS kernels so an operator runs on Apple Silicon, covering the native function dispatch as well as the host side of the operator. cuda-index-width helps decide whether a CUDA kernel needs 32-bit or 64-bit index math, the kind of decision that matters once a tensor is large enough to overflow a 32-bit index.

One level up sits the PyTorch 2 compiler stack. pt2-bug-basher debugs failures anywhere across it, from Dynamo graph breaks and Inductor codegen errors to AOTAutograd crashes and Triton kernel failures. aoti-debug narrows in on one path specifically, AOTInductor's ahead-of-time compilation, chasing segfaults, device mismatches, and constant-loading failures.

What handles documentation, CI, and everything else?

The remaining five skills round out the job outside the code itself. docstring writes docstrings for PyTorch functions and methods in the project's own conventions, and document-public-apis goes further, finding undocumented public APIs and wiring them into the project's Sphinx-based documentation. ci-metrics answers questions about PyTorch's own continuous integration, pulling from GitHub Actions, its HUD dashboard, and Grafana to explain job durations, failures, and queue times. pyrefly-type-coverage migrates a single file to PyTorch's stricter Pyrefly type checking, adding the annotations it requires for every function, class, and attribute in it. skill-writer is the exception: it is not about PyTorch's own code at all, but a general walkthrough for creating new Claude Code skills, kept in this repository alongside the rest.

Who should install this, and who shouldn't?

This is a maintainer and contributor plugin, not a machine learning one. All 16 components are skills; there are no agents, commands, MCP servers, or hooks in it, and every one of them assumes you have PyTorch's own source checked out and are changing something inside it: an operator, a kernel, an issue, or a pull request. If you use PyTorch as a dependency to build and train your own models, installing it gives you nothing to invoke. If you are working on PyTorch itself, whether that means a single operator fix or a deep dive into the compiler, most of what you would reach for by hand is already named above.

How do I install it?

Flowy indexes pytorch/pytorch straight from its public GitHub repository. Install it from the listing page linked at the top of this guide, and the 16 skills above become available in that session.