The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PRD documents which could contain malicious instructions designed to influence the audit outcome.
Ingestion points: Untrusted artifact paths (e.g., docs/02_PRD/...) and the document content itself.
Boundary markers: The skill implements a 'de-anchoring' strategy, explicitly instructing subagents to 'Disregard the author self-assessment score' and citing specific playbook checks to prevent anchoring or injection from the document's own metadata.
Capability inventory: File system read/write operations, Bash utility command execution, and invocation of downstream skills (e.g., doc-prd-fixer).
Sanitization: Mitigation is handled via explicit instructions to the model to ignore specific metadata fields (e.g., *_ready_score, audit_score) in the untrusted input.
[COMMAND_EXECUTION]: The skill executes Bash shell commands to manage directories and track execution time for its break-circuit policy.
Evidence: Use of mkdir -p to prepare blackboard directories and date +%s to implement a timeout mechanism.
These are utility commands used for lifecycle and state management within the agent framework and do not involve executing external or unvalidated remote code.

doc-prd-audit