The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted trace data from the Progress Observability Platform, which presents an indirect prompt injection surface.
Ingestion points: Raw trace content is retrieved via the get_observation_details_with_content tool in SKILL.md.
Boundary markers: The skill requires including a SECURITY_NOTE in generated prompts and uses specific delimiters for example blocks, as defined in references/frame.md.
Capability inventory: The skill uses read-only MCP tools and does not have the ability to write files or execute code, as noted in references/mcp.md.
Sanitization: The skill mandates redacting PII (emails, SSNs, credit cards) and defanging boundary tokens (e.g., <input>) in quoted content to prevent subversion of the prompt structure, detailed in references/frame.md.
[DATA_EXFILTRATION]: The skill retrieves observability data from mcp.observability.progress.com. This is a vendor-owned resource accessed via a read-only MCP interface for the purpose of generating context-aware evaluator prompts, as described in references/mcp.md.

generate-eval