The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes feature documentation and skill definitions from the local filesystem. Maliciously crafted content in these files could attempt to influence the agent's behavior, although the use of read-only subagents for this phase serves as a mitigation.
Ingestion points: SKILL.md and files within the features/ directory of the target skill (found in .agents/skills, .claude/skills, or .cursor/skills).
Boundary markers: No explicit delimiters are defined for the subagent prompts during the source wave phase.
Capability inventory: Execution of local helper scripts, file system writes (audit logs), and Git operations (opening PRs).
Sanitization: The skill does not explicitly sanitize the text content of the feature files before they are read by subagents.
[COMMAND_EXECUTION]: The skill identifies and executes local helper scripts (such as doctor or map-check) that are bundled with the target verification skill being maintained.
[EXTERNAL_DOWNLOADS]: The UPSTREAM.md file references the official Cursor plugins repository on GitHub for version tracking and update procedures. This reference to a well-known service is documented for provenance purposes.

maintain-verification-skill