The Agent Skills Directory

[SAFE]: The skill provides a structured framework for delegating tasks to specific AI model tiers (e.g., Sonnet for implementation, Opus for review) based on reasoning requirements.
[COMMAND_EXECUTION]: The skill includes shell commands for environment validation and configuration synchronization. These commands execute local scripts (validate.mjs and sync-config-dir.sh) located within the agent's standard configuration directory (~/.claude).
[INDIRECT_PROMPT_INJECTION]: The skill defines a routing layer that handles task data. It mitigates potential issues by instructing the agent to perform manual verification of subagent output (Step 3), requiring the agent to compare diffs against specifications and re-run tests rather than trusting subagent claims.
[DATA_EXPOSURE]: The included validation script (validate.mjs) reads local configuration files in ~/.claude/agents/ and ~/.claude/skills/ to verify model pinning, tool access, and conduct rules. This access is localized for configuration integrity and does not exfiltrate data.
[DYNAMIC_EXECUTION]: The skill uses a Node.js script (validate.mjs) to parse YAML frontmatter and tool arrays from agent definition files. The script performs static analysis and consistency checks on these files without executing their content or incorporating external input into execution paths.

claude-router