The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external URLs provided by the user or detected from the environment.
Ingestion points: Content is ingested through browser automation tools such as mcp__Claude_in_Chrome__read_page and mcp__Claude_in_Chrome__get_page_text (found in SKILL.md).
Boundary markers: There are no explicit instructions to the agent to disregard instructions embedded within the analyzed web content.
Capability inventory: The skill possesses powerful write capabilities, including Write, Edit, and Bash for git operations (git commit, git revert), which are used to modify the local codebase based on the audit results.
Sanitization: No explicit sanitization or filtering of the ingested web content is defined before it influences the agent's design critique and subsequent code modifications.
[COMMAND_EXECUTION]: The skill executes local scripts and shell commands to manage the state of the repository during the fix loop.
Evidence: It calls bash "$CLAUDE_PLUGIN_ROOT/scripts/clean-tree-check.sh" and bash "$CLAUDE_PLUGIN_ROOT/scripts/detect-base-branch.sh" to verify the working tree and branch context.
Context: These executions are scoped to the local environment and are used to ensure atomic commits for design fixes (git commit -m "style(design): ...").

design-review