The Agent Skills Directory

[SAFE]: The skill instructions and referenced prompts focus exclusively on prompt engineering QA, emphasizing fact-based reporting and rigorous test coverage without introducing security risks.
[SAFE]: No evidence of data exfiltration or credential harvesting was found. The skill explicitly mentions using dry runs and least privilege for sensitive operations.
[SAFE]: The skill does not perform any network operations, external downloads, or dynamic code execution. All logic is contained within the provided markdown and configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied prompts and datasets, creating a surface for potential indirect injection.
Ingestion points: The skill ingests user-supplied prompt versions, datasets, and input distributions as defined in prompts/prompt-testing.md.
Boundary markers: Instructions in SKILL.md require the agent to audit input and separate facts from assumptions.
Capability inventory: The skill contains no instructions for file writes, network requests, or subprocess execution.
Sanitization: The instructions recommend using masked data, mocks, and isolated environments for sensitive workflows.
[PRIVILEGE_ESCALATION]: The configuration in evals/local-rules.json explicitly restricts escalations by setting max_escalations to 0, further hardening the skill's execution environment.

prompt-testing