The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages a data ingestion surface that is susceptible to indirect prompt injection, though it includes significant logical guardrails to mitigate this risk.
Ingestion points: The skill processes untrusted data during Step 0 (Intake Triage), Step 1 (Evidence Ingestion), and when aggregating outputs from delegated subagents.
Boundary markers: The protocol utilizes specific tags like [NO-DATA] and [CONTRADICTION] to logically separate and highlight uncertain or conflicting external data within the decision memo.
Capability inventory: The agent has access to bash, view_file, write_to_file, and replace_file_content to perform its tasks.
Sanitization: The skill enforces a mandatory Socratic Adversarial Gate (requiring three counter-arguments) and a two-stage review process, which serves as a defensive filtration layer for all ingested data.
[COMMAND_EXECUTION]: The skill utilizes the system shell to perform cryptographic verification of proposed actions.
Evidence: The procedure in SKILL.md and the Staff Work Doctrine instructs the agent to use bash and sha256sum to generate hashes of execution payloads. This is a controlled and legitimate use of command execution intended to prevent unauthorized or accidental modifications to the filesystem.

secretary