The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses git and gh (GitHub CLI) to automate branch creation, Pull Request submission, and merge operations. It utilizes git worktrees in sibling directories to isolate changes and prevent interference with the user's main working environment.\n- [EXTERNAL_DOWNLOADS]: The skill executes bun install to resolve dependencies when a bun.lock file is present. This download occurs during the standard setup phase of the PR workflow.\n- [PROMPT_INJECTION]: Instructions include mandatory process rules (e.g., 'Always create a fresh git worktree', 'Do not merge until CI... passes') that enforce specific workflow constraints rather than attempting to subvert the agent's safety guidelines.\n- [INDIRECT_PROMPT_INJECTION]: The skill incorporates external feedback from Pull Request reviews to identify blocking issues.\n
Ingestion points: Feedback is ingested from GitHub API review comments (specifically from cubic-dev-ai[bot]) and task-based AI reviewer outputs.\n
Boundary markers: The skill checks for specific success indicators (e.g., 'No issues found') but lacks comprehensive isolation for the feedback body.\n
Capability inventory: The agent has the ability to modify code, execute shell commands, and push to remote repositories.\n
Sanitization: While approval is gated by string matching, the skill encourages acting on the content of external reviews to fix code issues, creating a potential vector for indirect instructions.

work-with-pr