The Agent Skills Directory

[SAFE]: The skill logic is implemented using standard Python libraries (re, dataclasses, enum). No dangerous functions such as eval(), exec(), or subprocess calls were identified in the source code.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it is designed to process and summarize untrusted contract text.
Ingestion points: The contract_text input variable in the extract_clauses and generate_summary methods (SKILL.md).
Boundary markers: The implementation lacks specific delimiters or system instructions to prevent the model from interpreting text within a contract as commands.
Capability inventory: The skill requests filesystem permissions (claw.json), which could be targetable via an injection.
Sanitization: Input text is processed for pattern matching but is not sanitized for malicious instructional content before inclusion in the final analysis report.
[EXTERNAL_DOWNLOADS]: The skill declares no external dependencies and does not perform any remote code fetching or package installations.
[DATA_EXFILTRATION]: No network-related commands (curl, wget, requests) or hardcoded credentials were found. All processing is performed on the provided input text strings.

contract-clause-extractor