The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local script scripts/validate_candidate.py to ensure data integrity. Technical analysis of the script confirms it performs static schema validation using standard Python libraries. It enforces strict type checking and format verification (e.g., regex for IDs and dates) without employing dynamic code execution or processing untrusted input in a way that allows for arbitrary command injection.
[DATA_EXPOSURE]: The skill references external URLs to official government entities (nrta.gov.cn) and established industry services (kuaishou.com, questmobile.com.cn, eastmoney.com). These references are used transparently for data sourcing and do not involve the exfiltration of sensitive system information or credentials.
[INDIRECT_PROMPT_INJECTION]: The skill processes external JSON/JSONL data files, representing a potential injection surface.
Ingestion points: Candidate files are ingested via the scripts/validate_candidate.py script and then processed by the agent.
Boundary markers: While formal delimiters are not explicitly defined in the prompt instructions, the skill mandates a human-in-the-loop gate in SKILL.md, requiring the user to explicitly approve content before it is committed.
Capability inventory: The skill utilizes local script execution for validation and a file-writing tool (create-data-context) for knowledge updates.
Sanitization: Strict schema validation is performed by the Python script, although it focuses on data structure rather than instruction sanitization. The requirement for explicit user approval significantly mitigates the risk of the agent obeying instructions embedded in the ingested data.
[PROMPT_INJECTION]: The skill contains highly restrictive internal logic and 'Inviolable Rules' that prevent unauthorized modifications to the knowledge base. These instructions serve as defensive constraints to ensure the agent only acts upon verified and user-approved information.

d-data-analysis-semantic-layer