tdg-personal/pytorch-patterns
PyTorch deep learning patterns and best practices for building robust, efficient, and reproducible training pipelines, model architectures, and data loading.
57
57%
Does it follow best practices?
Impact
Pending
No eval scenarios have been run
Passed
No known issues
Quality
Discovery
32%Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.
The description identifies the domain (PyTorch deep learning) and lists broad topic areas but lacks concrete actions and explicit trigger guidance. It reads more like a course title than a skill description, missing the 'Use when...' clause entirely and failing to enumerate specific capabilities that would help Claude select it appropriately.
Suggestions
Add an explicit 'Use when...' clause, e.g., 'Use when the user asks about PyTorch, neural network training, custom datasets, or building deep learning models in Python.'
Replace broad categories with specific concrete actions, e.g., 'Implements custom nn.Module architectures, configures DataLoaders with transforms, sets up distributed training, implements learning rate schedulers, and debugs gradient issues.'
Include more natural trigger terms and file/library references users would mention, such as 'torch', 'neural network', 'GPU training', 'CUDA', '.pt/.pth model files', 'optimizer', 'loss function'.
| Dimension | Reasoning | Score |
|---|---|---|
Specificity | Names the domain (PyTorch deep learning) and mentions some areas (training pipelines, model architectures, data loading), but these are broad categories rather than concrete actions like 'build custom DataLoaders' or 'implement learning rate scheduling'. No specific verbs describing discrete actions. | 2 / 3 |
Completeness | Describes what the skill covers at a high level but completely lacks any 'Use when...' clause or explicit trigger guidance. Per the rubric, a missing 'Use when...' clause should cap completeness at 2, and since the 'what' is also somewhat vague, this scores a 1. | 1 / 3 |
Trigger Term Quality | Includes relevant keywords like 'PyTorch', 'deep learning', 'training pipelines', 'model architectures', and 'data loading', which users might naturally say. However, it misses common variations like 'neural network', 'GPU training', 'DataLoader', 'loss function', 'optimizer', 'backpropagation', '.pt files', or 'torch'. | 2 / 3 |
Distinctiveness Conflict Risk | The mention of 'PyTorch' specifically helps distinguish it from general coding or other ML framework skills, but 'deep learning patterns and best practices' is broad enough to overlap with general ML skills, TensorFlow skills, or general Python coding skills. | 2 / 3 |
Total | 7 / 12 Passed |
Implementation
64%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
This is a solid reference of PyTorch best practices with excellent, executable code examples covering the most important patterns. Its main weaknesses are verbosity (the Good/Bad pattern adds bulk, and some explanations are unnecessary for Claude) and lack of progressive disclosure—everything is in one large file with no references to supplementary materials. The workflow dimension suffers from being a collection of independent patterns rather than a guided process.
Suggestions
Split advanced topics (mixed precision, gradient checkpointing, torch.compile, custom collate) into a separate ADVANCED.md file and reference it from the main skill
Remove explanatory comments that state the obvious to Claude (e.g., '# Crashes if no GPU', '# Dropout still active! BatchNorm uses batch stats!') to reduce token usage
Add a brief end-to-end workflow section showing the recommended sequence: seed setup → dataset → dataloader → model → training loop → validation → checkpointing, with explicit validation checkpoints between stages
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The skill is fairly comprehensive but includes some unnecessary commentary Claude already knows (e.g., 'Always set train mode', 'More efficient than zero_grad()', explaining what eval mode does). The Good/Bad pattern is useful but adds verbosity; some anti-patterns are well-known to Claude. At ~300 lines, it could be tightened significantly. | 2 / 3 |
Actionability | Every section provides fully executable, copy-paste-ready Python code with proper imports, type hints, and realistic function signatures. The code examples are complete and concrete, covering training loops, data loading, checkpointing, and optimization patterns. | 3 / 3 |
Workflow Clarity | While individual patterns are clear, there's no explicit end-to-end workflow showing how to sequence these components together (e.g., setup → data → model → train → validate → checkpoint). The training loop is well-structured but lacks validation checkpoints like 'verify GPU memory before proceeding' or 'check loss is decreasing before continuing training'. | 2 / 3 |
Progressive Disclosure | The content is a monolithic file with all patterns inline. For a skill this long (~300 lines), advanced topics like gradient checkpointing, custom collate functions, and mixed precision training could be split into referenced files. There are no cross-references to separate documents for deeper dives. | 2 / 3 |
Total | 9 / 12 Passed |
Validation
90%Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.
Validation — 10 / 11 Passed
Validation for skill structure
| Criteria | Description | Result |
|---|---|---|
frontmatter_unknown_keys | Unknown frontmatter key(s) found; consider removing or moving to metadata | Warning |
Total | 10 / 11 Passed | |
Reviewed
Table of Contents