huggingface - SKILL.md Agent Skill

name: huggingface description: Work with the Hugging Face Hub — search models/datasets, inspect model cards, download weights, and upload artifacts — via the huggingface-cli and the keyless public API. Token only needed for private/gated repos. Use when asked to "find a model for", "download this HF model", "search datasets", "what's on the Hub". allowed-tools: ["@coder", "Bash", "@webfetch"] triggers: - huggingface - hugging face - find a model for - download model - search datasets - model card - on the hub - baixar modelo - procurar modelo

Hugging Face Hub

Public Hub browsing is keyless (HTTP API); a token (HF_TOKEN) is only needed for private or gated repos and uploads.

Search (keyless)

curl -s "https://huggingface.co/api/models?search=whisper&sort=downloads&limit=10"
curl -s "https://huggingface.co/api/datasets?search=sentiment&limit=10"

Read id, downloads, likes, pipeline_tag.

Model / dataset card

@webfetch https://huggingface.co/<org>/<model>           # the card
curl -s "https://huggingface.co/api/models/<org>/<model>"  # metadata JSON

Download (huggingface-cli)

pip install -U "huggingface_hub[cli]"
huggingface-cli download <org>/<model> --local-dir ./model
# gated/private:
HF_TOKEN=... huggingface-cli download <org>/<model>

For a single file: huggingface-cli download <org>/<model> config.json.

Upload (needs token)

huggingface-cli login    # or set HF_TOKEN
huggingface-cli upload <org>/<repo> ./local-dir

Choosing a model

Search by task (pipeline_tag: text-generation, automatic-speech-recognition, etc.).
Sort by downloads/likes/recency; read the card for license, size, and intended use.
Report id, license, size, and a one-line fit assessment before downloading large weights.

Rules

Weights can be huge — confirm before downloading multi-GB repos; mention disk cost.
Respect model licenses; surface the license from the card.