The Agent Skills Directory

[SAFE]: A detailed security audit of the skill instructions and referenced documents found no evidence of malicious intent, obfuscation, or unauthorized access patterns.
[PROMPT_INJECTION]: The skill handles untrusted user data via the $ARGUMENTS parameter. 1. Ingestion points: User descriptions are captured in SKILL.md and wrapped in <feature description> tags. 2. Boundary markers: XML-style tags are employed to delimit user-provided content from system instructions. 3. Capability inventory: The skill utilizes internal agents for codebase review and document refinement, which are low-impact operations. 4. Sanitization: No explicit sanitization is performed, but the use of boundaries and the limited scope of available tools mitigate the risk of indirect injection.
[DATA_EXFILTRATION]: No network activity, hardcoded credentials, or access to sensitive file paths (e.g., .env, SSH keys) were detected. File system operations are strictly limited to writing brainstorm documents in the docs/brainstorm/ directory.

brainstorm