The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads machine learning datasets and checkpoints from official Hugging Face repositories and the Roboflow platform. It also references a third-party GitHub repository (abin24/Magnetic-tile-defect-datasets) to obtain specific training data for manufacturing defect inspection.
[COMMAND_EXECUTION]: The pipeline executes multiple bundled Python and Bash scripts (such as download_checkpoints.sh, prep_testcase.sh, and run_sdg.sh) to automate the end-to-end training and synthetic data generation workflow.
[PRIVILEGE_ESCALATION]: Documentation in references/docker.md provides instructions to use chmod 777 or sudo chown on host directories as a troubleshooting mechanism to resolve permission conflicts related to Docker bind mounts.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided defect specifications in JSONL format, specifically using the roi_prompt_defect_location field as an input to a vision-language model to guide the generation process.
Ingestion points: Training and inference datasets, as well as the defect_spec.jsonl configuration file.
Boundary markers: No specific delimiters or safety instructions are defined for the embedded prompts in the documentation.
Capability inventory: The skill can execute shell commands, Python scripts, and manage Docker containers with GPU access.
Sanitization: The provided references do not describe explicit sanitization or filtering for the natural language prompts ingested during the text2roi phase.
[DATA_EXFILTRATION]: The recommended Docker configuration mounts the host's /etc/passwd and /etc/group files as read-only volumes to allow the container to correctly resolve user and group identities for file ownership.

tao-generate-anomalies