The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the discovery and installation of third-party 'Agent Skills' from an external registry to extend the agent's capabilities based on the project stack. This is performed using the npx skills add command.
Evidence: Found in modes/tool-skills.md under Step 3, which details the installation of selected skills.
Mitigation: The skill implements a mandatory consent gate (TOOL-CONSENT) defined in modes/tool-skills.md, requiring explicit user approval before any external searches or installations are initiated.
[COMMAND_EXECUTION]: The skill executes various shell commands to perform its auditing functions, including git for repository analysis and initialization, as well as npx for skill management.
Evidence: Mentioned in SKILL.md (Portability section), modes/greenfield.md (Git integration), and modes/tool-skills.md.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted data from the project codebase to generate AI-readable documentation (AGENTS.md). Malicious instructions embedded in source code comments or manifests could potentially influence the resulting context files.
Ingestion points: The skill reads the entire project source code, manifest files (e.g., package.json), and git logs across multiple phases (modes/whole-repo.md, modes/area.md, modes/gapfill.md).
Boundary markers: The skill appends a 'Drafted by /audit' signature to all generated AGENTS.md files to explicitly signal to both humans and other agents that the content was tool-generated and requires verification.
Capability inventory: The skill possesses Bash, Write, Edit, and Agent tools, allowing it to modify the file system and spawn subagents based on the data it processes.
Sanitization: The instructions emphasize 'gap-filling' conservatively and prohibit overwriting curated prose, which acts as a secondary check against malicious modifications to existing documentation.

audit