The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection by ingesting and summarizing untrusted tabular data provided by the user.
Ingestion points: The skill loads user-provided datasets as X and Y (or XTrain/YTrain) in Step 1.
Capability inventory: The agent has the ability to write files (save_selected_models.m, export_workflow_script.m) and execute MATLAB code via the MCP server.
Boundary markers: While the skill uses structured steps, it lacks strict delimiters or explicit 'ignore instructions' warnings when the agent summarizes data characteristics in the Step 12 summary.
Sanitization: The references/dataprep.md workflow includes several data cleaning steps (omitting features, merging classes) which mitigate accidental data-driven behavior.
[DYNAMIC_EXECUTION]: The skill uses dynamic code generation and execution to provide its core functionality.
Script generation: scripts/helpers/export_workflow_script.m programmatically assembles a MATLAB .m script based on the current session's state and writes it to the file system to enable the user to reproduce the training workflow.
Dynamic invocation: scripts/helpers/build_model_definitions.m and scripts/helpers/retrain_with_hpo.m use feval() to call MATLAB functions determined at runtime by data characteristics.
Chained script execution: scripts/helpers/build_model_definitions.m uses run() to execute a series of catalog scripts (classifier_branches.m, branch_*.m) to assemble the list of candidate models.
[OBFUSCATION]: The skill includes pre-compiled MATLAB P-code files which are opaque to static analysis.
Evidence: scripts/helpers/count_categorical_levels.p and scripts/helpers/stratified_subsample.p. While these are standard pre-compiled formats for MATLAB intellectual property protection, they prevent full verification of the logic they contain. These are provided as bundled utility helpers by the vendor.

matlab-classify-tabular-data