The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements proactive defenses against both direct and indirect prompt injection. All subagent instruction files (such as subagents/adversary-seat.md and subagents/chair-seat.md) include explicit directives to treat content within the <decision_packet> tags as data only, to ignore any embedded imperative commands, and to report such instances as manipulation risks. Furthermore, the skill defines a clear ingestion surface for untrusted data while mitigating risk through structural delimiters and strictly enforced subagent independence. Specifically regarding the injection surface:
Ingestion points: The DECISION_SUBJECT, STATED_CLAIM, CONSTRAINTS, and CURRENT_LEAN inputs defined in SKILL.md.
Boundary markers: The orchestrator wraps user-provided and retrieved content in <decision_packet> tags to maintain separation from instructions.
Capability inventory: The skill performs file system writes to HANDOFF_PATH and optionally utilizes web search tools to verify claims.
Sanitization: Individual seat agents are instructed to ignore embedded instructions and prioritize independent analysis over provided payload imperatives, effectively sanitizing the input during the reasoning phase.
[DATA_EXFILTRATION]: The skill manages its output securely by using a deterministic slugification process for file naming and a collision policy that prevents overwriting existing files. File operations are limited to the resulting analysis handoff, and the use of research tools is strictly scoped to a verified evidence tier with required locators.

council-of-advisors