The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the host.exec() API in resources/api-reference.md, which allows for the execution of shell commands. The documentation provides examples such as git log, find, and npm commands, which are intended for developer automation workflows.\n- [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process various file formats (PDF, CSV, DOCX, XLSX) which are often untrusted external inputs. This creates a surface for indirect prompt injection attacks. However, the skill provides extensive documentation on mitigations:\n
Ingestion points: Data enters the context via parsers for PDF, CSV, DOCX, and XLSX as detailed in resources/api-reference.md.\n
Boundary markers: The resources/patterns.md file advises on explicit trust boundary framing and using the system.safety prompt to instruct the model to ignore instructions within data.\n
Capability inventory: The framework provides powerful capabilities including host.exec(), host.writeFile(), and host.readFile(), which could be abused if an injection is successful.\n
Sanitization: Mitigation patterns in resources/patterns.md recommend using strict defSchema definitions with additionalProperties: false to restrict the model's output and validating all tool arguments before use.

genaiscript