The Agent Skills Directory

[SAFE]: No malicious patterns, obfuscation, or unauthorized access attempts were identified in the skill instructions or methodology.
[PROMPT_INJECTION]: The skill design includes an indirect prompt injection surface because its primary purpose is to process and evaluate external prompts.
Ingestion points: The Subagent invocation contract section in SKILL.md and SKILL-ja.md handles the target prompt to be tested.
Boundary markers: The instruction template uses Markdown headers (e.g., ## Target prompt) to separate untrusted prompt content from the agent's evaluation instructions.
Capability inventory: The subagents are granted tool access to perform realistic scenario tests as defined by the user.
Sanitization: No sanitization is performed on the target prompts, as the skill is intended to evaluate the raw, potentially ambiguous behavior of the agent when encountering them.
[COMMAND_EXECUTION]: While the skill workflow involves running subagents that may execute commands to complete their assigned scenarios, this behavior is contained within the standard agent task orchestration and is the intended functionality of the evaluation workflow.

empirical-prompt-tuning