The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a workflow that fetches content from external, untrusted sources (HuggingFace model cards) to extract metadata like descriptions and licenses. This represents an indirect prompt injection surface as the agent is instructed to process and interpolate this data into local documentation files without explicit sanitization or boundary markers.
Ingestion points: SKILL.md (Phase 1, Step 1) instructs the agent to fetch HuggingFace repository pages to extract model information.
Boundary markers: The process lacks specific instructions to treat fetched content as data rather than instructions, potentially allowing a malicious model card to influence agent behavior.
Capability inventory: The skill utilizes file-writing capabilities (JSX, MDX), source control commands (git), and GitHub management tools (gh), which could be misused if an injection occurs.
Sanitization: There are no specified sanitization steps for the fetched strings before they are written to the repository.
[COMMAND_EXECUTION]: The skill workflow involves executing several local system commands, including git for branch management, gh for pull request creation, and Mintlify tools (mint validate, mint dev) for documentation testing.
Logo Generation: The skill includes instructions to create a Python virtual environment in /tmp to install the Pillow library and run a script to process brand logos provided by the user.
Verification Scripts: It calls project-specific scripts like node docs/scripts/check_cookbook_configs.mjs to validate configuration tags.
[EXTERNAL_DOWNLOADS]: The skill fetches external data and dependencies as part of its normal operation.
Package Installation: Instructions include installing Pillow via pip and sglang[diffusion] via uv.
External Content: It retrieves model metadata from HuggingFace URLs to populate cookbook templates.

cookbook-add-model