The Agent Skills Directory

[SAFE]: Tool access is limited to a secure subset of shell commands (ls, cp, mkdir, git) and local file operations via the frontmatter configuration.
[SAFE]: Explicit safety instructions prevent data deletion and mandate the creation of backup copies in a legacy folder before any changes.
[SAFE]: The skill provides instructions to restrict the agent to the project directory and its subdirectories, mitigating directory traversal risks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project files (Ingestion points: Read, Glob, Grep in SKILL.md) which could contain embedded instructions. However, boundary markers and sanitization are absent, but risks are mitigated by the lack of code execution capabilities and limited tool access (Capability inventory: Write, Edit, git).

code-archaeology