The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is specifically designed to handle untrusted input from GitHub PRs (code, comments, descriptions). It implements multiple defensive layers:
Boundary Markers: Explicitly instructs agents via a 'UNTRUSTED_CONTENT' block to treat all PR data as information to be analyzed, not instructions to be followed.
Tampering Detection: Instructs subagents to identify and report 'review-process tampering'—attempts within the PR content to influence the review outcome or the agent's behavior.
Policy Enforcement: Review rules are fetched from the 'base' commit SHA, preventing a PR from modifying the guidelines it is being judged by.
Skeptic Verification: Uses an adversarial 'skeptic' subagent to independently verify findings and filter out fabrications or pre-existing issues.
[EXTERNAL_DOWNLOADS]: The skill uses the standard gh CLI tool to interact with the GitHub API. All network operations are directed towards GitHub's official services, which is expected for a PR review tool.
[REMOTE_CODE_EXECUTION]: There is no evidence of arbitrary code execution. The skill performs static analysis of the PR diff and does not execute the code it reviews unless specifically requested by the user for a 'focused local check'.
[DATA_EXFILTRATION]: No data exfiltration patterns were detected. Interaction is limited to the user and the GitHub repository being reviewed.
[CREDENTIALS_UNSAFE]: No hardcoded credentials were found. The skill relies on the existing authentication of the gh tool.

github-review-pr