The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill performs an 'Agentic Review' of Simulink models where the LLM evaluates semantic guidelines based on model data.
Ingestion points: Model metadata, block names, and parameter values are ingested from .slx files using model_read and model_query_params as described in references/uncheckable-guidelines-review.md.
Boundary markers: The skill uses reporting templates and note blocks to separate analysis, but does not employ strict delimiters for raw model data during the evaluation process.
Capability inventory: The skill has high-privilege capabilities including model_edit (to modify model structure and parameters) and evaluate_matlab_code (to execute arbitrary MATLAB scripts).
Sanitization: No explicit sanitization or escaping of model-derived strings is mentioned. The risk is partially mitigated by instructions requiring user confirmation before applying any changes in 'Fix Mode'.
[COMMAND_EXECUTION]: The skill uses evaluate_matlab_code to run local MATLAB scripts located in the scripts/ directory. These scripts (model_advisor_run.p, etc.) are P-coded (MATLAB's protected format), which is standard for MathWorks-distributed tools but limits transparency into the exact command execution logic.

checking-model-compliance