Hugging Face Skills is an out-of-the-box toolkit that contains documentation, scripts and tools that allow AI agents to create datasets, train models, run evaluations, manage tasks, and publish papers. It works seamlessly with tools such as Claude Code, OpenAI Codex, Gemini CLI, Cursor, etc. -just install it with a simple command and invoke the skill in the command, such as “Training Skills with HF Models.” The tool automatically performs complex Hugging Face Hub operations, saving time and allowing agents to complete tasks accurately without having to write code manually.
If you are still manually copying and pasting Hugging Face API documentation, or awkwardly explaining “How to help me run an SFT fine-tuning task” to the AI in Cursor/Claude, you must read this article.
Recently, Hugging Face quietly updated a new one called skills warehouse. At first glance, the name was plain, but on closer inspection, I called him “good guy”-this is simply a “plug-in cram class” for AI Agents.
What are Hugging Face Skills?
Simply put, it is not a library for you to use, but for your hand. AI Assistant (Cursor, Claude Code, Windsurf, Gemini CLI) A prepared “skill book”.
In the past, when you said to the AI,”Help me pass this dataset to HF.” The AI might have to go through the document for a long time and finally write a string of code that might report errors.
Now, you only need to install this “skill” and AI will master it instantlyComplete standard operating procedures (SOPs)。It knows how to slice data most efficiently, how to select graphics cards based on video memory, and how to automatically send an email informing you that you are finished training.
What pain points can it solve?
As an ML engineer or developer, the most annoying thing is not the model logic, but theOperation of crushing into slag：
- Environment configuration: This library depends on that library, and it collapses when the version changes.
- Memory anxiety: Choose A100 or H100? Is there enough memory memory to perform this fine-tuning?
- Process fault: After practicing the model, you need to write an evaluation report, and after writing it, you need to send the paper to arXiv, and finally send it to Hub.
huggingface/skills Take all these thingsstandardizationYes. It turns tasks that originally required writing hundreds of lines of code into “atomization capabilities” that AI assistants can directly call.
A few operations that make you call “It’s cool”
1. “Help me practice a model quickly!”
by huggingface-llm-trainer Skills, you only need one sentence: “Use SFT methods to fine-tune this model.” AI will automatically check your hardware and help you write the most robust scripts. Even high-end postures such as GRPO and DPO have been trained into muscle memory.
2. “Paper publishing is a one-stop process”
This is what I think is the most outrageous. it has a huggingface-paper-publisher Skills.
What can it do? From generating Markdown templates and organizing BibTeX, to automatically associating models and datasets, and even helping you sync your papers to arXiv. AI becomes your scientific assistant, and you are only responsible for producing Ideas.
3. “Data Set Porter”
No need to download it locally. call huggingface-datasets Skills, let AI help you filter, clean, and analyze directly in the cloud. You just need to command in the dialog box: “Help me see how much dirty data there is in this dataset” and it will get busy on its own.
Use gracefully?
The most exciting thing about this project is that itSeamless adaptationThe most popular AI programming tool now:
- Claude Code： run
/plugin install hf-cli@huggingface/skillsDone. - Cursor： mating
.cursorrulesAI instantly transformed into HF experts. - Gemini CLI： Even the configuration file has been written for you.
“Standardized plug-ins” in the Agent Era
If LLM is the brain, then Hugging Face Skills is a library of high-order algorithms pre-installed in the brain.
It represents a trend:Future developers are no longer “people who write code” but “people who schedule skills.” We no longer need to worry about how to call the specific API, but focus on “what problems I want to solve.”
The project is not yet fully popular, so while there are not many Star, go to GitHub to collect a handful of wool and try installing it in your Cursor first. Trust me, the tacit understanding of “AI knows which graphics card to choose next before you even ask” is really addictive.
Github：https://github.com/huggingface/skills
Oil tubing: