The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from Parquet and CSV files and incorporates it into a Markdown report.
Ingestion points: External data is ingested via gaps_parquet and source_pool_csv files as described in SKILL.md.
Boundary markers: There are no explicit delimiters or instructions to ignore embedded content when interpolating data into the Routing_Report.md output.
Capability inventory: The skill possesses file system read/write capabilities and shell command execution via the Bash tool.
Sanitization: Input labels are converted to uppercase strings for comparison, but no further sanitization is performed on the data before it is written to output reports.
[DYNAMIC_EXECUTION]: The skill relies on generating and executing code at runtime for its core functionality and validation.
The SKILL.md provides Python recipes that the agent is instructed to execute via Bash to filter data and generate summaries.
Multiple validation scripts in the hooks/ directory (e.g., routing-artifacts-check.sh and routing-coverage-check.sh) use Python heredocs to execute dynamic analysis code on the generated artifacts.
[DATA_EXFILTRATION]: The packaging mechanism accesses session-specific metadata and logs.
The hooks/routing-package.sh script is designed to access session transcripts and project logs from ~/.claude/projects/ to include them in the skill's output directory.
While this is part of the intended SDA iteration packaging, it involves the movement of session data into generalized output folders.

tao-route-visual-changenet-samples