The Agent Skills Directory

[Interaction with Untrusted Data]: The skill is designed to ingest and process content from GitHub pull requests, including body text, commit messages, and diffs, which constitute external and untrusted data.
Ingestion points: The skill uses gh pr view, gh api, and gh pr diff in SKILL.md to gather pull request context.
Boundary markers: The instructions mandate a clear "write boundary" and a two-phase process requiring explicit user confirmation before any mutations occur.
Capability inventory: Includes the ability to post comments and reviews via gh api and gh pr comment.
Sanitization: The skill specifies using temporary files or direct API input to ensure content is handled literally without shell substitution.
Context: This interaction is inherent to the skill's purpose. The risk of indirect prompt injection is mitigated by the requirement for a human-in-the-loop confirmation gate ("Post these comments? and stop") before any feedback is submitted to GitHub.
[Command Execution via GitHub CLI]: The skill uses the GitHub CLI (gh) to perform read and write operations scoped to a specific repository.
Evidence: Commands such as gh pr view <PR> --repo docker/docs and gh api repos/docker/docs/pulls/<PR>/reviews are used throughout SKILL.md.
Context: These operations are restricted to the official documentation repository. The skill's design emphasizes that the agent should only mutate state (e.g., posting comments) after a user has reviewed and approved the specific draft text, which reduces the risk of unauthorized or unintended command execution.

review-pr