The Agent Skills Directory

[External References]: The skill references external resources hosted on official Microsoft GitHub repositories (github.com/microsoft/ai-agent-eval-scenario-library and github.com/microsoft/triage-and-improvement-playbook). These links point to documentation and scenario frameworks intended to support the evaluation process. These references originate from the skill's authoring organization.
[Processing of External Data]: The skill is designed to ingest and analyze evaluation results provided by the user, including test cases, sample inputs, and agent responses. This represents a potential surface for indirect prompt injection if the test data contains malicious instructions. However, the skill acts as a diagnostic aid for the user, and the risk is mitigated by its structured reporting format and focus on interpretation rather than automated execution of the content.
Ingestion points: Step 1 (Gather Eval Results) asks the user to provide failing test cases and manifest metadata.
Boundary markers: The workflow uses a structured triage report format, which helps in separating data from instructions.
Capability inventory: The skill is primarily instructional and report-generating; it does not exhibit capabilities for command execution or network exfiltration.
Sanitization: The skill relies on standard LLM handling of user-provided context for analysis purposes.

eval-triage-and-improvement