The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from pull request comments and review threads. This attack surface is well-managed through a strictly defined "Bounded discussion loop" that requires explicit user agreement on classifications and actions for every item. Ingestion points: Step 3 in SKILL.md collects review feedback via the gh CLI. Boundary markers: Step 5 mandates an evidence-led discussion loop where the user must agree on every decision and Step 7 requires final user approval before any remote actions. Capability inventory: The skill uses the Bash tool for git operations, local file modifications, and GitHub interaction. Sanitization: Step 8 explicitly requires using body files for GitHub replies to prevent Markdown content from being misinterpreted as shell commands.
[COMMAND_EXECUTION]: The skill executes repository-defined commands (e.g., from Makefile or README.md) during the pre-flight phase to establish a healthy baseline. It also manages code changes through focused git commits and pushes. It explicitly forbids destructive commands like git reset --hard and uses safe data handling practices for external content.
[EXTERNAL_DOWNLOADS]: The skill communicates with GitHub's official services via the gh CLI to fetch pull request metadata and post responses. These interactions target a trusted service provider and are scoped to the active pull request.

pr-review-grill