The Agent Skills Directory

[COMMAND_EXECUTION]: The skill involves the execution of local development tools and scripts to validate test specifications. This includes using dotnet run to execute the repository's internal SkillValidator project, performing dotnet build to verify test fixtures, and running python scripts to check evaluation quality. These operations are restricted to local paths within the project structure (such as the eng/ directory) and are consistent with the skill's primary function as a testing and scaffolding tool.
[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting existing skill and agent documentation to generate new evaluation scenarios, which presents a potential surface for indirect instruction influence.
Ingestion points: The skill reads existing skill content from SKILL.md or .agent.md files located in the plugins/ directory to understand the target's capabilities (detailed in the 'Workflow' section).
Boundary markers: The instructions explicitly require the agent to generate stimuli and rubrics that avoid using the target skill's specific vocabulary or terminology, which serves as a delimiter to prevent overfitting and accidental instruction mimicry.
Capability inventory: Capabilities include executing local C# projects (dotnet run), running shell commands for environment setup and fixture validation, and executing local Python scripts for quality checks.
Sanitization: All generated evaluation files are subjected to structural and quality checks through the skill-validator tool and the check_eval_quality.py script, which enforce schema compliance and detect structural defects before the specifications are accepted.

create-skill-test