The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local project repository and interpolates it into instructions for a review subagent.
Ingestion points: Reads build-plan.md, project-overview.md, architecture.md, design.md, and code diffs via git diff (referenced in review-agent-prompt.md and modes/review.md).
Boundary markers: The content is placed under markdown headers in the review-agent-prompt.md template but lacks strict delimiters or "ignore embedded instructions" warnings for the interpolated data.
Capability inventory: The review subagent is granted the Bash, Read, Write, and Agent tools (as described in modes/review.md).
Sanitization: Project file contents are inlined into the subagent prompt without validation, escaping, or filtering.
[COMMAND_EXECUTION]: The skill's verification mode executes commands specified in the project's own configuration files.
Evidence: modes/verify.md uses the Bash tool to run docker compose, package.json scripts, and Makefiles found within the analyzed repository.
[DYNAMIC_EXECUTION]: The skill dynamically generates subagent prompts and executes project-specific scripts based on runtime detection.
Evidence: modes/review.md populates a subagent template (review-agent-prompt.md) with dynamic content from the file system. modes/verify.md identifies and executes local build tools based on project metadata.

dev-check