The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted sensor and historical production data through its Python modules.
Ingestion points: predict_failure, train_model, and sync_with_physical functions in SKILL.md and record_production in MANUFACTURING_EXECUTION_SYSTEM_MES.md accept external dictionary and DataFrame objects.
Boundary markers: None explicitly defined to separate data from instructions within the logic, but the data is strictly parsed as numerical features or specific schema types.
Capability inventory: The skill possesses Read, Write, and Edit tool permissions, but the scripts do not execute shell commands, perform network operations, or access sensitive system files.
Sanitization: Input data is processed using standard data science libraries (numpy, pandas, sklearn) for feature extraction and statistical calculations, which effectively sanitizes input against prompt-based injection by converting it to numeric types before processing.
[EXTERNAL_DOWNLOADS]: The skill references several external resources for manufacturing standards (ISA-95, OPC UA, MTConnect). All linked domains are official industry organization websites and are used for documentation purposes only.
[DYNAMIC_EXECUTION]: The skill uses scikit-learn to train a RandomForestClassifier at runtime. This model training is performed on locally provided historical data and does not involve the execution of arbitrary code or untrusted deserialization.

manufacturing-expert