The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external data including engineering plans, adversarial findings, and repository policies which could potentially contain malicious instructions.\n- Ingestion points: Processes plans, findings, and policies as described in SKILL.md.\n- Boundary markers: The skill requires using a "fresh session" for the review, providing isolation from the primary execution context.\n- Capability inventory: The instructions explicitly restrict the agent to read-only operations, forbidding plan rewrites, code implementation, or spawning additional workers in SKILL.md.\n- Sanitization: While no explicit sanitization is mentioned, the strict read-only and no-execution constraints significantly mitigate the impact of potential indirect injections.

simplicity-review