The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill requests a Hugging Face access token with write scope to perform authenticated repository reads and Space publication.
Evidence: The skill checks for existing tokens via huggingface_hub.get_token() and, if unavailable, prompts the user to provide one. The token is stored as a Space secret (HF_TOKEN) on the destination repository, which is a standard security practice for Hugging Face Spaces.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted metadata from Hugging Face model cards and README files to determine parameters for code generation and UI layout.
Ingestion points: huggingface_hub.ModelCard.load and HfApi().list_repo_files in SKILL.md.
Boundary markers: Absent; untrusted metadata is processed directly without explicit delimiter separation.
Capability inventory: Generates and writes Python code (app.py), requirements files, and markdown files; performs repository creation and secret management via HfApi; executes testing via the gradio CLI.
Sanitization: Present; Phase 4 enforces a mandatory batched approval step where the user reviews the content of all three generated files (app.py, requirements.txt, README.md) before any publication occurs.
[PROMPT_INJECTION]: A static detector identified potential concealment logic regarding the automation of trigger words.
Evidence: The instructions in references/adapting-to-the-lora.md suggest automatically prepending model trigger words to user prompts. This is a standard user experience enhancement for model demos and does not represent a malicious attempt to hide agent instructions from the user.
[EXTERNAL_DOWNLOADS]: The skill references and downloads dependencies from official and well-known repositories within the Hugging Face and Lightricks ecosystems.
Evidence: Generated configurations pull from github.com/huggingface/diffusers and github.com/Lightricks/LTX-Video for specialized model support.
[COMMAND_EXECUTION]: The skill utilizes the gradio CLI to verify the functionality of the published application.
Evidence: Phase 6 uses gradio info and gradio predict to smoke-test the remote Space endpoint once it is running.

huggingface-lora-space-builder