The Agent Skills Directory

[PROMPT_INJECTION]: The file tests/samples/injected-claude.md contains an instruction to ignore previous instructions. This is a dedicated test fixture used to verify that the auditing tool's safety detection band correctly identifies and flags concealed directives.
[COMMAND_EXECUTION]: The auditing tool (scripts/claude_check.py) uses subprocess.run() to call git log and git rev-list. These calls are used exclusively to gather metadata about the repository's commit history to identify stale memory files. The test suite also executes the tool via subprocess for verification.
[DYNAMIC_EXECUTION]: The skill uses importlib and __import__ to load its analysis engine and discover tests. Specifically, scripts/_bootstrap.py and scripts/claude_check.py dynamically resolve and load the sibling rabbit-writes library, while tests/run.py dynamically loads test modules.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process CLAUDE.md and AGENTS.md files, which are untrusted data sources.
- Ingestion points: Content is read from local markdown files in the repository root or submodules.
- Boundary markers: No explicit delimiters are used for the audited content.
- Capability inventory: The skill can read/write files (write is gated on user approval) and execute local Git commands.
- Sanitization: The skill proactively mitigates this risk by including a specialized 'safety' analysis band designed to detect concealed text or instructions aimed at the agent within the audited files.

rabbit-claude-md