The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates a workflow that ingests data from local machine learning training checkpoints and environment configuration files.
Ingestion points: Reads model weights (.pt, .pth), configuration files (opt.yaml), and .env files from the local filesystem (SKILL.md).
Boundary markers: The instructions explicitly warn against using insecure methods like source .env for pre-existing files, which could execute malicious shell code, and instead recommend using python-dotenv for safe, scoped reading.
Capability inventory: Includes file system read/write access, shell execution (pip, uv), and network operations through the roboflow SDK (SKILL.md).
Sanitization: Leverages built-in validation within the Roboflow SDK to verify model types, architectures, and sizes before the upload is finalized.
[COMMAND_EXECUTION]: The skill involves the use of shell commands for standard developer environment setup and deployment tasks.
Instructs the use of uv venv and pip install for dependency management and python for executing the deployment logic.
[EXTERNAL_DOWNLOADS]: Recommends the installation of official software libraries for interacting with the Roboflow platform.
Fetches dependencies such as roboflow and rfdetr from the Python Package Index (PyPI), which are standard vendor-owned resources.

roboflow-custom-weights-upload