The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several git commands such as git status, git diff, git show, and git merge-base to establish review context and inspect code changes. It also instructs subagents to use git worktree add to create temporary directories for revision inspection, which involves active filesystem operations.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a surface for indirect prompt injection as it processes data from the repository that may be under the control of an external actor (e.g., in a PR or downloaded plan).
Ingestion points: Implementation descriptions, requirement files, and git diff outputs are interpolated directly into the subagent instructions in references/code-reviewer.md.
Boundary markers: While the template uses markdown headers for organization, it does not employ robust delimiters or explicit instruction overrides to protect against malicious instructions embedded within the codebase or plan files.
Capability inventory: The skill possesses the capability to execute shell commands (git), read local documentation (style guides), spawn subagents, and perform file writes when applying authorized fixes to the codebase.
Sanitization: There is no evidence of sanitization, filtering, or validation of the data extracted from the repository before it is passed to the reviewer subagent.
[DYNAMIC_EXECUTION]: The skill's behavior is partially defined by workspace-local configuration files, specifically .claude/CLAUDE.md. The instructions mandate following this file to determine the review orchestration (e.g., number of reviewers and model selection), which allows for dynamic modification of the skill's logic based on the environment's current state.

code-review