The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an automated workflow that processes external markdown review reports to generate structured modifications. This creates a surface for indirect prompt injection where adversarial content in a review report could influence the agent's behavior.
Ingestion points: sdd parse-review reads external markdown files (e.g., reports/review.md).
Boundary markers: No explicit delimiters are used to wrap the review content during parsing to distinguish data from instructions.
Capability inventory: Subprocess execution through sdd verify and file modification via sdd apply-modifications.
Sanitization: The workflow incorporates a --dry-run preview and post-application schema validation, though these rely on user/agent review to catch logical injection.
[COMMAND_EXECUTION]: The modification schema explicitly supports a command field for verification steps. While intended for running tests (e.g., pytest), this capability allows for the definition and subsequent execution of arbitrary shell commands in the local environment if a modification file is malicious.

sdd-modify