The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data including git diffs, commit messages, and CODEOWNERS files through its subagents. To mitigate risks, it implements a comprehensive safety chain:
Ingestion points: subagents/diff-analyzer.md and subagents/pr-drafter.md read repository content and git output.
Boundary markers: The skill provides explicit instructions to treat all analyzed repository content strictly as data and to report any imperative text (e.g., instructions to skip steps or change reviewers) as suspected injection.
Capability inventory: The skill utilizes git push and platform CLI commands (gh, glab) to create pull requests, but limits these to plain pushes without force flags.
Sanitization: It employs a rigorous multi-stage approval process where a preview is frozen, digested, and verified against the final submission to prevent tampering.
[EXTERNAL_DOWNLOADS]: The skill is configured to fetch technical documentation and syntax guidance from established, trusted organizations such as GitHub, GitLab, and Google. These fetches are limited to one URL per decision and are used strictly as background reference for command flags and writing standards, not for executable scripts.

pr-creator