The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted user prompts as its primary operational data.
Ingestion points: The Analysis phase in SKILL.md takes a 'target prompt' for evaluation and optimization.
Boundary markers: The instructions require the agent to 'Preserve the original prompt verbatim' and keep metadata like output_mode outside the target prompt, which helps distinguish input from instructions but does not provide complete isolation.
Capability inventory: The skill has no dangerous capabilities; it does not utilize filesystem writes, network operations, or subprocess execution in any of its files (SKILL.md, references/*.yaml, agents/*.yaml).
Sanitization: No explicit sanitization or escaping of the target prompt is performed before interpolation into the analysis phases, relying on the underlying LLM's safety filters.

prompt-optimization