The Agent Skills Directory

[COMMAND_EXECUTION]: The reference file references/doc-drift-patterns.md advises the agent to verify documentation by running the commands it contains: "Check setup/build/test commands actually run — the fastest drift check is executing them." This instructs the agent to execute potentially untrusted code from files like CLAUDE.md or README.md without providing a mechanism for sanitization or safe execution.
[EXTERNAL_DOWNLOADS]: The SKILL.md documentation contains instructions for installing external skills via a script (python scripts/install-external-skills.py). This establishes a dependency on external code that is downloaded and executed on the host system.
[PROMPT_INJECTION]: The skill processes external documentation files which could contain hidden or indirect prompt injections. The advisory to execute commands found in these files increases the risk of the agent obeying malicious instructions embedded in the audited content.
Ingestion points: Document files (README.md, CLAUDE.md, etc.) in the audited repository.
Boundary markers: None present in the instructions.
Capability inventory: Shell command execution (suggested for verification), file reads, and grep.
Sanitization: No sanitization or validation of the audited content is described.

documentation-drift