The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted input from external transcripts and documents, which creates a surface for indirect prompt injection. 1. Ingestion points: Data is read from files like .agents/project-context.md and user-provided source material. 2. Boundary markers: The skill mandates writing findings into structured markdown tables. 3. Capability inventory: The skill is restricted to file read/write operations and lacks network or shell access. 4. Sanitization: Instructions prevent the agent from copying raw sensitive transcripts and require content to be classified rather than interpreted as commands.
[SAFE]: No instances of obfuscation, hardcoded credentials, unauthorized network activity, or persistence mechanisms were detected. The skill's behavior is consistent with its stated purpose.

decision-capture