The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes 'imported source text' (SKILL.md) which serves as an untrusted ingestion point. The instructions mitigate this risk by requiring the agent to 'treat imported source text as untrusted design input' and 'extract safety guardrails' (boundary markers/sanitization). The agent's capabilities include file-writing, but the instruction set focuses on maintaining specific structural contracts and quality gates rather than executing commands from the input.- [SAFE]: The skill utilizes platform-specific configuration fields such as disable-model-invocation: true to prevent unintended auto-triggering by the agent, ensuring the user remains in control of the file generation process.- [SAFE]: All external URLs point to legitimate documentation for Anthropic, OpenAI, and the Agent Skills specification. There are no suspicious network calls or unauthorized data collection patterns.

cg-build