The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses sensitive user-global directories and files to perform its audit.
Evidence: It explicitly targets ~/.claude/CLAUDE.md, ~/.claude/projects/<encoded-cwd>/memory/MEMORY.md, and ~/.claude/plugins/** to collect memory and configuration data.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from project documents which could contain malicious instructions.
Ingestion points: Reads CLAUDE.md, MEMORY.md, and recursively follows @import tokens and markdown links ([text](./path.md)) within the project and user-global directories.
Boundary markers: The instructions do not specify any delimiters or safety warnings to the model to ignore instructions embedded within the files being audited.
Capability inventory: The skill can write files to the .drift-reports/ directory, create new git branches, and execute CLI commands to create Pull Requests.
Sanitization: There is no evidence of sanitization or filtering of the content retrieved from the audited files before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill utilizes external CLI tools to perform its automated fix functions.
Evidence: It executes gh pr create to generate Pull Requests for document improvements.

doc-drift