How to use the LoRA Space Builder agent skill

What is an agent skill, and what does this one do?

An agent skill is a folder of instructions and helper files that an AI agent can load when a request matches the skill's description. This one is named huggingface-lora-space-builder and comes from the huggingface/skills collection, under the Apache-2.0 licence.

A LoRA is a small add-on file that changes how a larger image or video model behaves. When you ask, the agent reads the LoRA's files and model card on the Hugging Face Hub, picks the matching base diffusion pipeline, and writes a small web app with a UI tailored to what that LoRA does. It then publishes that app as a Hugging Face Space, private by default, so you can try it in a browser.

The skill covers base models such as Qwen-Image, Qwen-Image-Edit, LTX-Video, Wan, FLUX and SDXL, and tasks including text-to-image, image-to-image, text-to-video, image-to-video and video-to-video. It also tries to respect the model card's recommendations for trigger words, steps, guidance, LoRA scale and example inputs.

Install the skill (ZIP download or Claude apps upload)

Download the skill ZIP from the skill's page on this directory. You have two options depending on how you use your agent.

For Claude Code, unzip the folder into your personal skills directory ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only inside that project. The skill folder should contain SKILL.md along with its references folder.

For the Claude apps, open Settings, go to Capabilities, then Skills, and upload the skill there.

• Personal (all projects): ~/.claude/skills/
• Single project: .claude/skills/ inside the project folder
• Claude apps: Settings → Capabilities → Skills

What you need before you start

The skill needs Python with huggingface_hub and diffusers, plus Gradio for the generated app. Its scripts run and it uses the network, so your agent environment must allow both.

You need a Hugging Face account and an access token with write scope. The token is needed to publish the Space and to read private or gated repos. Create one at https://huggingface.co/settings/tokens.

The Space runs on GPU hardware. ZeroGPU is the default target. The source does not list prices for GPU hardware, so check Hugging Face's own pricing before you publish.

• Python with huggingface_hub and diffusers
• Gradio (used by the generated app)
• Hugging Face account and write-scope access token
• GPU hardware on Hugging Face Spaces (ZeroGPU by default)

Trigger the skill with a natural request

You do not need a special command. Describe what you want in plain language that mentions a Space, demo, Gradio app or playground, and name the LoRA repo. The skill's description lists trigger phrases such as create, generate, ship or publish a Space, demo, Gradio app or playground for a LoRA. It also triggers when you say you trained or host a LoRA on the Hub and want to share it.

Include the repo ID if you have it, for example username/my-lora. The agent will then read the repo first. If the repo is public, it proceeds without a token. If it is private or gated, the agent checks whether you are already logged in before asking you for a token.

• Make a Gradio demo Space for my LoRA at username/my-lora
• Publish a playground for my Qwen-Image LoRA so my team can try it
• I trained a FLUX LoRA and want to share it as a browser demo
• Build a video-to-video demo for my LTX LoRA, private by default

How the agent builds the demo

The agent first gathers information from the LoRA's files and model card. It then picks the base pipeline and inference recipe, designs a UI for the LoRA's task and inputs, and writes app.py, requirements.txt and README.md together. You are shown all three files for one batched approval before anything is published.

Good output has only the controls the LoRA needs, reports the seed used and shows a clear error when input is missing. The source says the agent should trust the model card's parameters such as steps, guidance, CFG, LoRA scale and dtype, and it prefers loading the weights with pipe.load_lora_weights.

• Reads the LoRA repo files and model card
• Chooses the matching base pipeline
• Designs a UI for the specific task and inputs
• Shows app.py, requirements.txt and README.md for approval
• Publishes the Space as private

Safety, privacy and what to check first

Publishing writes to your Hugging Face account using a write-scope token, so only give the token to an agent you trust and that is running in an environment you control. Treat the token as a password.

Do not use this skill for a private or unreleased LoRA until you have checked the privacy settings. A demo can receive your prompts and media, and weights in a public-facing Space could expose them. The skill publishes as a private Space by default, but confirm that setting before sharing the link.

If you only want to run a LoRA locally inside a script, this skill is the wrong tool, because it is built around publishing a Space. The skill also does not cover training a LoRA or general web apps.

• Use a write-scope token only in a trusted environment
• Confirm the Space is private before sharing it
• Check privacy before sending unreleased weights or prompts
• Not for local scripts, training or general web apps

Tips for better results

Give the agent the details the model card might lack. The skill asks you once, in a single batched message, if it cannot infer the task or what the LoRA does. A one-line answer such as relight portraits or extend videos to wider aspect ratios helps it design a better UI.

Ask for the LoRA's recommended settings to be respected, and point out any trigger word you know of. Review the three files before approving, especially the chosen base model, since the source warns that the wrong pipeline can load but produce wrong output.

• Tell the agent the sub-task, such as relighting, face swap or style, not just the task type
• Name the trigger word if the card does not show it
• Check the chosen base model before approving the files