The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because it ingests untrusted data from a 'Reconnaissance Dossier' and interpolates it into subagent prompts without sufficient boundary markers or sanitization.
Ingestion points: The reconnaissance dossier is read in SKILL.md (Phase A) and passed as input to the subagents in proposer-prompt.md and devils-advocate-prompt.md.
Boundary markers: The skill uses markdown headers for structure but lacks explicit delimiters or instructions to the subagents to ignore embedded commands within the ingested data.
Capability inventory: The skill dispatches subagents via the Task tool and interacts with the user via the question tool to reach a final architectural consensus.
Sanitization: There is no evidence of escaping or filtering logic applied to the external reconnaissance content before it is processed by the agents.

the-grilling