ml-docs

name	ml-docs
description	On-demand ML/data science library expert. Use when the user asks how to use any function, class, or method from NumPy, Pandas, scikit-learn, Matplotlib, TensorFlow, Keras, PyTorch, Seaborn, SciPy, statsmodels, XGBoost, LightGBM, Hugging Face Transformers, OpenCV, NLTK, spaCy, Plotly, Dask, PySpark, SQLAlchemy, or Jupyter. Fetches and synthesizes official API docs, parameter reference, and working code examples. Also use when the user asks "how do I do X in pandas/numpy/torch/sklearn", needs to understand a deep learning layer or training loop, asks about NLP pipelines, computer vision transforms, statistical tests, SQL ORM patterns, or big data ops.
allowed-tools	Bash(uv run * scripts/process.py *) Read Write WebFetch Glob Grep
argument-hint	[library] [topic or function name]
model	haiku
effort	low
compatibility	>=1.0
metadata	{"category":"reference","tags":["docs","api-reference","numpy","pandas","pytorch","sklearn","huggingface","xgboost","lightgbm","statsmodels"],"phase":"any"}

ML Docs

Context: $ARGUMENTS

Quick start

• Look up a specific function/class: → Step 1, then Step 2
• Understand a concept or workflow: → Step 1, pick a topic URL, fetch it
• Compare options across libraries: → run Step 1 for each library, fetch in parallel

When to use

• User asks how to use any function, class, or method from a supported library
• User asks "what parameters does X take" or "what does Y return"
• User needs working code examples for a task (groupby, cross-validation, fine-tuning, etc.)
• User asks about a concept: broadcasting, autograd, attention, sparse matrices, etc.
• User wants to know which library/function to use for a task
• User hits an error and needs to check expected behavior from the official docs

Step 1 — Resolve the documentation URL

Run process.py resolve to get the prioritized list of URLs to fetch:

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve \
  --library <library-name-or-alias> \
  --query "<function, class, or topic>"

The script returns a JSON object with fetch_in_order — a list of URLs ranked by specificity. Start with priority 1. If it returns 404 or empty content, move to priority 2, then 3.

Library aliases accepted: See references/guide.md for the full alias table.

Examples:

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library pandas --query "DataFrame.groupby"
uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library torch --query "autograd"
uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library sklearn --query "cross_val_score"
uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library transformers --query "pipeline"
uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library scipy --query "hypothesis testing"

To list all supported libraries:

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py list

Step 2 — Fetch and synthesize

Use WebFetch on the URLs returned by Step 1, in order. Stop at the first URL that contains useful content. Synthesize into a direct answer including:

1. What it does — one sentence
2. Signature — function/class signature with parameter types if available
3. Key parameters — name, type, default, what it controls
4. Return value — type and meaning
5. Working example — copy-pasteable, minimal, correct
6. Common gotchas — anything non-obvious from the docs (deprecations, defaults that surprise)

Do not paste raw documentation. Synthesize it.

Step 3 — Cross-library questions

When the question spans multiple libraries (e.g. "how do I use a PyTorch model with scikit-learn's cross-validation"), run Step 1 for each library in parallel, fetch both, then synthesize a combined answer that shows how they integrate.

Gotchas

• Keras vs tf.keras: Keras 3.x (keras.io) is standalone and backend-agnostic. tf.keras is the older TF-bundled version. If the user has TF < 2.16, they likely have tf.keras. Check which one they're importing before advising.

• PyTorch torch. vs torch.nn.functional: Many operations exist in both places with different call conventions. torch.nn.Conv2d is a module (stateful), F.conv2d is a function. The docs are on different pages — resolve for the right one.

• scikit-learn class paths: API URLs use the full dotted path, e.g. sklearn.linear_model.LogisticRegression, not just LogisticRegression. Include the module prefix in the --query argument for direct API lookups.

• Pandas 2.x breaking changes: DataFrame.append is removed in 2.0. df.swaplevel behavior changed. If the user shows old code, check the 2.0 migration guide: https://pandas.pydata.org/docs/whatsnew/v2.0.0.html

• Hugging Face pipeline task names: They changed between versions. Always fetch the current docs rather than recalling task names from memory (e.g. "text-generation" vs "text2text-generation").

• spaCy model names: en_core_web_sm/md/lg/trf are not installed by default. The docs show nlp = spacy.load("en_core_web_sm") but users need python -m spacy download en_core_web_sm first. Always mention this.

• OpenCV Python bindings: The Python docs at docs.opencv.org are C++ first. Prefer fetching the Python tutorials (/tutorial_py_*) over the raw C++ API pages.

• PySpark version: API paths differ between Spark 3.x and older versions. The default URL targets latest. If the user specifies a version, adjust the URL.

• Dask DataFrame is NOT Pandas: Dask DataFrames don't support all Pandas operations. Always check the Dask API index rather than assuming Pandas parity.

• LightGBM vs XGBoost parameter names: They use different names for the same concept (e.g. num_leaves in LightGBM vs max_leaves in XGBoost). When helping with both, always fetch both sets of parameter docs.

Examples

Example 1: Look up a Pandas function

User: "How do I use pandas pivot_table?"

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library pandas --query "DataFrame.pivot_table"

Fetch priority-1 URL → synthesize parameters, example, gotchas about aggfunc defaulting to mean.

Example 2: PyTorch training loop

User: "How does autograd work in PyTorch?"

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library torch --query "autograd"

Fetch the autograd topic page → explain requires_grad, .backward(), zero_grad(), with example.

Example 3: Hugging Face fine-tuning

User: "How do I fine-tune BERT with the Trainer API?"

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library transformers --query "trainer"
uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library transformers --query "bert"

Fetch both in parallel → synthesize a complete fine-tuning workflow.

Example 4: Statistical test

User: "How do I run a t-test in Python?"

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py resolve --library scipy --query "hypothesis testing"

Fetch stats topic page → show scipy.stats.ttest_ind and ttest_rel with example and interpretation.

Example 5: Unknown library scope

User: "How do I do X?"

Run:

uv run ${CLAUDE_SKILL_DIR}/scripts/process.py list

Scan the list for the relevant library. If the task spans multiple (e.g. plot a pandas DataFrame with Seaborn), resolve for both, fetch in parallel, synthesize combined answer.

Reference docs

Read references/guide.md for:

• Full alias table for all 19 libraries
• Direct topic URL index (skip the script for common topics)
• Library selection guidance (which library to use for a given task)