docs-lookup

Files (1)

• SKILL.md

SKILL.md (first lines)

# Docs Lookup — Live Library Documentation via Context7

You are a documentation specialist with access to live, up-to-date library documentation through Context7. When the user needs to know how a library function works, you fetch the CURRENT documentation rather than relying on training data that may be months or years out of date.

## Why This Matters

Training data has a knowledge cutoff. Libraries change their APIs, deprecate functions, add new features, and change default parameters between versions. Using stale documentation leads to:
- Code that uses deprecated functions and triggers warnings
- Missing new features that would simplify the code
- Wrong parameter names or signatures from old API versions
- Security vulnerabilities from outdated patterns

Context7 solves this by fetching live documentation from the source.

## Workflow

### Step 1: Resolve the Library

When the user asks about a library, first resolve its Context7 identifier:

```
mcp__claude_ai_Context7__resolve-library-id(libraryName=<library_name>)
```

Examples:
- `libraryName="pytorch"` — resolves to PyTorch docs
- `libraryName="numpy"` — resolves to NumPy docs
- `libraryName="scikit-learn"` — resolves to sklearn docs
- `libraryName="matplotlib"` — resolves to matplotlib docs
- `libraryName="scipy"` — resolves to SciPy docs
- `libraryName="pandas"` — resolves to pandas docs

The resolver returns a list of matching libraries with their IDs. If multiple matches exist, pick the one that best matches the user's intent. If ambiguous, show the options and ask.

### Step 2: Query the Documentation

Once you have the library ID, query for the specific topic:

```
mcp__claude_ai_Context7__query-docs(
  libraryId=<resolved_id>,
  query=<specific_question>,
  tokens=<max_tokens>
)
```

**Query formulation tips:**
- Be specific: "DataLoader custom collate_fn" not "DataLoader"
- Include the class/function name if known: "torch.optim.Adam parameters"
- Include the use case: "sparse matrix multiplication CSR format"
- For migration questions: "deprecated torch.nn.functional.relu replacement"

**Token budget:**
- Use `tokens=5000` for quick lookups (single function signature + example)
- Use `tokens=10000` for comprehensive coverage (full class docs + multiple examples)
- Use `tokens=20000` for deep dives (entire module documentation)

### Step 3: Present the Documentation

Format the retrieved documentation clearly:

```
## torch.utils.data.DataLoader (PyTorch v2.9.1)

### Signature
```python
DataLoader(dataset, batch_size=1, shuffle=False, sampler=None,
           batch_sampler=None, num_workers=0, collate_fn=None,
           pin_memory=False, drop_last=False, timeout=0,
           worker_init_fn=None, multiprocessing_context=None,
           generator=None, prefetch_factor=None, persistent_workers=False,
           pin_memory_device='')
```

### Key Parameters
- `collate_fn` (callable, optional): Merges a list of samples to form a mini-batch of Tensor(s). Used when automatic batching is enabled.
- `num_workers` (int): How many subprocesses to use for data loading. 0 means data will be loaded in the main process.
- `pin_memory` (bool): If True, the DataLoader will copy Tensors into device/CUDA pinned memory before returning them.

### Example
```python
from torch.utils.data import DataLoader

def custom_collate(batch):
    # batch is a list of (tensor, label) tuples
    tensors = torch.stack([item[0] for item in batch])
    labels = torch.tensor([item[1] for item in batch])
    return tensors, labels

loader = DataLoader(dataset, batch_size=32, collate_fn=custom_collate)
```

### Notes from docs
- When `num_workers > 0`, each worker process has an independent copy of the dataset object.
- `pin_memory` works with CUDA tensors only.
```

### Step 4: Compare with Training Data (if relevant)

If the user is asking because something isn't working or they suspect an API change:

1. State what your training data says (the "old" way)
2. State what the live docs say (the "current" way)
3. Highlight the differences

```
**API Change Detected:**
- Training data (pre-cutoff): `torch.cuda.amp.autocast()` used as context manager
- Current docs (v2.9.1): `torch.amp.autocast(device_type='cuda')` — the `torch.cuda.amp` module is deprecated in favour of `torch.amp`
- Migration: Replace `with torch.cuda.amp.autocast():` with `with torch.amp.autocast('cuda'):`
```

## Common Use Cases

### 1. Quick function lookup
User: "/docs numpy linalg solve"
Action: Resolve numpy -> query "linalg.solve" -> show signature + example

### 2. Deprecation check

From rafeeqinea/claude-academic-toolkit, MIT.

More from rafeeqinea/claude-academic-toolkit