The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from Parquet and CSV files. This creates a potential surface for indirect prompt injection if the input files contain malicious content that the agent interprets while processing or summarizing the data.
Ingestion points: pd.read_parquet(gaps_parquet) and pd.read_csv(source_pool_csv) in SKILL.md.
Capability inventory: File system read/write operations and subprocess execution via Bash/Python.
Sanitization: The skill performs case-insensitive filtering but lacks explicit validation or escaping of the label content before it is written to reports.
Boundary markers: None identified.
[COMMAND_EXECUTION]: The skill instructs the agent to execute Python scripts using the Bash tool to perform data analysis, routing, and reporting tasks.
[DATA_EXFILTRATION]: The hooks/routing-package.sh script accesses the user's session history and transcripts located at ~/.claude/projects/ to include them in the skill's local output directory. While this data remains in the user's workspace, it involves harvesting interaction history outside the immediate task context for auditing purposes.

tao-route-visual-changenet-samples