The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes external data including git logs and GitHub pull request comments (ingested in SKILL.md), creating a surface for potential indirect prompt injection. To maintain safety, the skill includes explicit boundary markers and instructions to treat all imported data as untrusted and never as instructions. Its capability inventory includes tools like git, gh, mktemp, and file-writing capabilities (for CLAUDE.md and skill files). These are protected by a strict sanitization contract for data returned by the repo-scout agent, which uses regex validation for counts, paths, IDs, and links to ensure data integrity and prevent command injection.
Command Execution: The skill utilizes shell-based utilities such as git, gh, mktemp, and rmdir to export repository metadata and manage temporary working directories. These operations are scoped to repository analysis and include cleanup procedures for generated temporary files. The instructions specifically forbid the use of untrusted data from the repository within any shell commands.
External Data Access: The skill interacts with GitHub's API via the gh tool to retrieve pull request comments. This is a standard operation for repository analysis tools and targets a well-known service to perform its intended function. The skill includes human-in-the-loop checkpoints, requiring explicit champion consent before reading from these sources.

patterns