The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (CSV and Excel files) to generate profiling reports and cleaning scripts.
Ingestion points: Data enters the agent's context through the Read, Grep, and Glob tools when analyzing user-provided datasets in Stage 1 and Stage 2.
Boundary markers: The instructions in SKILL.md include an 'Anti-Hallucination' section and a 'DATA PRIVACY WARNING' suggesting deidentification. However, there are no structural delimiters defined to isolate untrusted data content (such as column headers or cell values) from the agent's internal instructions during the automated code generation process.
Capability inventory: The skill is authorized to use Bash, Write, and Edit, and is explicitly designed to generate Python or R code that is then executed in the local environment.
Sanitization: While the skill instructs the model not to 'fabricate' variable names, there is no explicit instruction to sanitize or escape data-derived strings before interpolating them into generated scripts or Markdown tables.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to run diagnostic scripts (e.g., scripts/check_reverse_coding.py) and to execute the final cleaning code. Although the workflow requires user approval for each cleaning action, the execution of generated code remains a primary capability that could be targeted by sophisticated injections.

clean-data