The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes a 'launch prompt' that can contain blocks for audit scope, tool exclusions, and user emphasis. This input influences the orchestration logic and report prioritization, which creates a potential surface for indirect prompt injection if the launch prompt is provided by an untrusted source.
Ingestion points: Phase 0 of SKILL.md describes parsing the launch prompt for DIFF SCOPE, EXCLUSIONS, and USER EMPHASIS blocks.
Boundary markers: The skill relies on specific text headers to delimit configuration data within the prompt.
Capability inventory: The skill uses the Agent tool to launch sub-tasks, performs file system operations (read/write in scratch/), and executes project-wide searches via Grep as defined in SKILL.md.
Sanitization: The instructions do not describe any mechanisms to sanitize or validate the freeform text provided in the USER EMPHASIS block.

axiom-health-check