The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on external data by ingesting files and diffs provided by the caller or retrieved from a git repository.
Ingestion points: The skill uses "caller's files or diff" or the "current diff against the base branch" as the primary scope for the subagent analysis (SKILL.md).
Boundary markers: The instructions provide specific defensive logic to separate content from instruction, such as requiring the agent to "Reject application-code edits" and "scope escapes" in the subagent's report.
Capability inventory: The skill is capable of spawning subagents via spawn Task and performing file deletions or edits based on the subagent's findings.
Sanitization: The skill implements a verification step where the agent must audit suppressions, check "MUST KILL" reasons, and verify "IMPORTANT" or "do not remove" markers using /how or /why commands before acting on them.

no-comments