The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references several research papers and documentation sources from trusted and well-known organizations including Microsoft, OpenAI, and Vercel. These references are included for informational context and the skill explicitly states that they do not need to be fetched during execution.
[COMMAND_EXECUTION]: To perform its primary function of auditing, the skill instructs the agent to execute repository-native commands (e.g., build scripts, tests). This is protected by a set of 'Hard rules' that mandate safe execution, read-only defaults, and prohibit destructive actions or the deployment of code.
[DATA_EXPOSURE]: The instructions include defensive measures to prevent data leakage, such as explicitly forbidding the inclusion of secrets in reports and treating shared-environment credentials as out-of-bounds.
[PROMPT_INJECTION]: As a tool designed to ingest and process data from untrusted repositories (READMEs, manifests, etc.), it presents an attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through the inspection of the target repository's files and command outputs (documented in references/repo-inspection.md).
Boundary markers: The skill establishes clear boundaries through its operating modes (Audit vs Improve) and strict execution rules defined in SKILL.md.
Capability inventory: The agent is authorized to run shell commands and modify files, but these are scoped to the development journey and gated by human review guidelines.
Sanitization: The skill mandates the filtering of secrets and warns against 'vibes-only' findings, requiring all reports to be based on concrete repository artifacts.

dx-first