The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to ingest external data from a 'Reconnaissance Dossier' and interpolate it into prompts for various subagents (Proposer, Devils-Advocate, Synthesizer). This architectural pattern creates an indirect prompt injection surface where malicious instructions inside a dossier could influence subagent behavior.
Ingestion points: The 'Reconnaissance Dossier' and 'Project Constitution' files provide the content for subagent tasks.
Boundary markers: Absent; the prompt templates in proposer-prompt.md, devils-advocate-prompt.md, and synthesizer-prompt.md do not use specific delimiters or instructions to ignore embedded commands in the dossiers.
Capability inventory: The orchestrator dispatches subagents via a Task tool and interacts with the user via a question tool; it does not directly execute shell commands or perform network operations outside the platform's subagent framework.
Sanitization: No explicit sanitization of the input data is performed before it is passed to the subagents.
[SAFE]: The skill operates within the expected bounds of a multi-agent orchestration workflow. It does not attempt to access sensitive system files, establish persistence, or exfiltrate data. The logic is focused on managing a debate process with hard round limits and clear user checkpoints.

the-grilling