The Agent Skills Directory

[SAFE]: The core instructions in SKILL.md define a structured routing and writing workflow. It emphasizes reader outcomes and structural verification rather than attempting to override agent safety protocols or system instructions.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface as it is designed to read, review, and update existing documents in the workspace. While this is an inherent vector for indirect prompt injection if those files contain malicious instructions, the skill's workflow mitigates this by requiring strict adherence to document-type templates and structural validators.
[DATA_EXFILTRATION]: No suspicious network activity was identified. URL references within the evaluation files (such as in google-play.md) point exclusively to legitimate first-party documentation domains like support.google.com. Local file paths observed in the evaluation logs (e.g., /Users/buyonglee/...) appear to be artifacts of the author's testing environment and do not constitute a threat to the end-user.
[COMMAND_EXECUTION]: While the evaluation harness (evals/README.md) mentions the execution of local Python and Node.js scripts for validation (e.g., run_production_evals.py, validate_package.py), these are development-time tools for the skill author. No arbitrary or dangerous shell command execution was found within the skill's runtime instructions for the AI agent.

document-writing