Hugging Face Skills are standardized definitions for AI/ML tasks, such as dataset creation, model training, and evaluation, designed for interoperability with major coding agent tools including Anthropic's Claude Code, OpenAI Codex, and Google DeepMind's Gemini CLI. The core methodology centers on a self-contained folder structure for each skill, packaging instructions, scripts, and resources.

Each skill's primary definitional component is a SKILL.md file, which includes YAML frontmatter specifying the skill's name and description, followed by detailed guidance for the AI agent when the skill is active. This SKILL.md serves as the primary instruction set. While "Skills" is a term adopted from Anthropic, the repository ensures compatibility across different agent ecosystems. For instance, OpenAI Codex utilizes an AGENTS.md file for agent instructions, and Google Gemini employs a gemini-extension.json file. This repository intelligently adapts, generating the necessary files (e.g., via scripts/generate_agents.py) to conform to each agent's expected format, allowing a single skill definition to function across multiple platforms. A fallback agents/AGENTS.md is provided for agents without direct skill support.

Integration with various agent tools is facilitated through specific installation procedures:

• Claude Code: Users register the repository as a plugin marketplace via /plugin marketplace add huggingface/skills, then install individual skills using $/plugin install <skill-name>@huggingface/skills$.
• OpenAI Codex: Codex automatically identifies skills through the generated AGENTS.md file. Users can verify loaded instructions with codex --ask-for-approval never "Summarize the current instructions.".
• Google Gemini CLI: Integration is achieved by installing the provided gemini-extension.json file either locally (gemini extensions install . --consent) or directly from the GitHub URL (gemini extensions install https://github.com/huggingface/skills.git --consent).

The repository includes a set of pre-defined skills, each addressing a specific Hugging Face ecosystem functionality:

• hugging-face-cli: Facilitates Hugging Face Hub operations, including model/dataset downloads, file uploads, repository management, and cloud compute job execution.
• hugging-face-datasets: Manages Hugging Face Hub datasets, supporting repository initialization, configuration/system prompt definition, streaming row updates, and SQL-based querying/transformation.
• hugging-face-evaluation: Aids in managing evaluation results within Hugging Face model cards, covering extraction from READMEs, import from external APIs, and custom evaluations with frameworks like vLLM/lighteval.
• hugging-face-jobs: Enables running and managing compute jobs on Hugging Face infrastructure, including Python script execution, scheduled job management, and status monitoring.
• hugging-face-model-trainer: Focuses on training and fine-tuning language models using TRL on Hugging Face Jobs infrastructure, supporting methods like SFT, DPO, GRPO, and reward modeling, alongside GGUF conversion, hardware selection, cost estimation, and Trackio monitoring.
• hugging-face-paper-publisher: Manages research paper publication on Hugging Face Hub, supporting page creation, linking to models/datasets, authorship claims, and markdown article generation.
• hugging-face-tool-builder: Assists in building reusable scripts for chaining Hugging Face API calls or automating repetitive tasks.
• hugging-face-trackio: Provides functionality for tracking and visualizing ML training experiments, including metric logging via Python API and retrieval via CLI, with real-time dashboards synchronized to HF Spaces.

Once a skill is installed, users activate it by directly mentioning it in their coding agent instructions, for example, "Use the HF LLM trainer skill to estimate the GPU memory needed for a 70B model run." The agent then automatically loads the corresponding SKILL.md instructions and any associated helper scripts.

Skill contribution and customization involve copying an existing skill folder, renaming it, and updating its SKILL.md frontmatter with a new name and description. Supporting scripts, templates, and documents referenced by the instructions can be added or edited. A concise, human-readable description must be added to .claude-plugin/marketplace.json, distinct from the SKILL.md description which guides agent activation. The structure is validated by running python scripts/generate_agents.py, followed by reinstalling or reloading the skill bundle in the coding agent. The marketplace.json file plays a crucial role in presenting skills to human users in a plugin marketplace, ensuring consistency between skill names/paths in SKILL.md files and the marketplace listings.