The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input via the $ARGUMENTS variable and interpolates it directly into the prompts for several subagents, including data-engineer, data-scientist, and ml-engineer in SKILL.md.
Ingestion points: User input through the $ARGUMENTS placeholder is distributed into multiple subagent prompts across all four phases of the pipeline.
Boundary markers: No delimiters or explicit instructions to ignore embedded commands are used when interpolating $ARGUMENTS into the prompts.
Capability inventory: The subagents are granted broad capabilities, including generating "complete implementation code," "automation scripts," "Kubernetes manifests," and "monitoring configuration" (e.g., in SKILL.md phases 1 through 4).
Sanitization: There is no evidence of validation, filtering, or escaping performed on the user input before it reaches the subagents.
[EXTERNAL_DOWNLOADS]: The skill references a wide array of well-known and trusted third-party technologies and frameworks as part of its MLOps workflow.
Evidence: Mentions of tools such as DVC, lakeFS, Feast, Tecton, MLflow, Weights & Biases, Optuna, Ray Tune, Horovod, FastAPI, TorchServe, Kubeflow, Airflow, ArgoCD, Terraform, Helm, and Prometheus. These are standard industry tools for machine learning operations and are used here within their intended primary purpose.

machine-learning-ops-ml-pipeline