How do Skills relate to AGENTS.md and gemini-extension.json?

[Hugging Face Skills](https://github.com/huggingface/skills) keeps the capability inside skill folders, where SKILL.md carries guidance and guardrails, while AGENTS.md and gemini-extension.json act as tool-specific entrypoints and compatibility adapters. Think of it as: folders hold content and executable assets, entry files define distribution and loading. This keeps skills versioned in one structure while remaining recognizable across multiple agent tools.

How do you make skills governance rollback-safe for a team?

Treat each skill folder as a releasable unit: stable structure, indexable YAML metadata, and scripts/templates versioned together. Use normal code review to control interface changes and run small regression tasks to validate outputs and script behavior before merging. If a change drifts, rollback is just a version revert, restoring automation stability.

Why is it a stronger #alternative-to-agents-md organization pattern?

AGENTS.md is closer to repo-level global instructions, great for general rules but weak for bundling task scripts and templates, and not ideal for per-task install/reuse. Skills make the unit a folder, so prompts, scripts, templates, and resources ship as a task bundle and you can install only what you need. For real engineering workflows, that granularity improves reuse, auditability, and clear boundaries.

Hugging Face Skills Deep Dive: Portable Agent Skills Repo

Pain Points vs Innovation

✕Traditional Pain Points	✓Innovative Solutions
Agent rules often sprawl: some live in README, some in AGENTS.md, others in tool-specific settings, ending up as non-reviewable and non-revertible instruction fragments.	Hugging Face Skills treats a skill as a task bundle: SKILL.md plus scripts/templates ships the “how” as an installable folder, not scattered text.
Prompts alone can’t reliably encode real workflows: dataset creation, training, evaluation, and publishing need scripts, templates, and guardrails to prevent drift.	It uses ACP for shared task definitions and reuses the same skills across multiple agent tools via Claude Code, Codex, and Gemini CLI entrypoints.

Deployment Guide

1. Clone the repo and inspect available skill folders

bash

1git clone https://github.com/huggingface/skills.git && cd skills && ls

2. Register the repo as a Claude Code plugin marketplace

bash

1/plugin marketplace add huggingface/skills

3. Install a skill by folder (example: training skill)

bash

1/plugin install hf-llm-trainer@huggingface-skills

4. Mention the skill in your agent request to activate it

bash

1Use the HF LLM trainer skill to estimate GPU memory for a 70B run.

Use Cases

Core Scene	Target Audience	Solution	Outcome
Dataset Template Factory	Data and Labeling Teams	Use a dataset-creator skill to generate structured samples, prompt templates, and checks	Standardize data and reduce rework
Training Pipeline Bootstrap	Training Engineers	Use an llm-trainer skill to estimate cost/VRAM and scaffold training scripts	Faster time-to-train with fewer config mistakes
Evaluation and Reporting Automation	MLOps Teams	Use a model-evaluation skill to orchestrate eval jobs and produce comparison reports	Repeatable quality gates with traceable metrics

Hugging Face Skills

What is it?

Pain Points vs Innovation

Architecture Deep Dive

Deployment Guide

1. Clone the repo and inspect available skill folders

2. Register the repo as a Claude Code plugin marketplace

3. Install a skill by folder (example: training skill)

4. Mention the skill in your agent request to activate it

Use Cases

Limitations & Gotchas

Frequently Asked Questions