The Agent Skills Directory

[SAFE]: The skill operates as a structured review process using markdown and YAML artifacts. Analysis of the provided files reveals no evidence of command execution, network exfiltration, or persistence mechanisms. The skill relies on static references and human-in-the-loop validation.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external artifacts which may contain content from the web, creating a surface for indirect prompt injection.
Ingestion points: The agent ingests search outputs and vocabulary maps provided in the task context, as described in SKILL.md.
Boundary markers: The skill uses condition C20 in references/conditions.md as a boundary marker, explicitly instructing the agent to ignore any instructions found in fetched content.
Capability inventory: The skill has no capabilities for network operations, file writing, or subprocess execution; its scope is limited to text analysis and reasoning.
Sanitization: The skill relies on the LLM's adherence to the safety instructions provided in the conditions file to prevent the execution of embedded instructions.
[EXTERNAL_DOWNLOADS]: The calibration fixtures reference documentation from well-known technology services for context and calibration.
Evidence: Fixture files include URLs from Microsoft (Visual Studio Code), Google (Chrome Web Store), Shopify, and Figma to serve as examples of valid data sources. These references are documented neutrally and are part of the intended calibration functionality.

reviewing-platform-ecosystem-prior-art-survey