The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a consolidated bash script (PHASE_A_SCRIPT.md) to gather structural information about the project. This script performs file existence checks, directory inspections, and line count operations to inform the audit process without modifying the system state.
[DATA_EXPOSURE]: The skill reads project configuration and documentation files, including CLAUDE.md, package.json, and various settings files (settings.json, .claude/settings.json, settings.local.json). These files are accessed solely to assess the project's health and readiness for Source Driven Development (SDD). The results are saved to a local report file at .claude/audit-report.md and are not transmitted to external servers.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it ingests and processes untrusted data from the project filesystem.
Ingestion points: Project files including CLAUDE.md, ai-context/*.md, package.json, pyproject.toml, skill definitions in .claude/skills/, and project documentation files.
Boundary markers: No explicit delimiters are used when reading content from these files.
Capability inventory: File system read access, local file write access for report generation, and bash execution for discovery purposes.
Sanitization: The skill performs basic structural validation (e.g., verifying YAML manifest syntax) but does not escape or sanitize content before processing. However, the risk is considered low given the diagnostic, read-only purpose of the tool.

project-audit