The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell and Node.js scripts (such as grill-brief.mjs, grill-answers.mjs, and _resolve-plugin-root.sh) to perform business logic extraction and verdict processing.
[DYNAMIC_EXECUTION]: Script paths are computed at runtime using environment variables (CLAUDE_PLUGIN_ROOT) or by searching the local plugin cache ($HOME/.claude/plugins/cache/*/unwind/*/). Sourcing shell scripts (source) and executing Node.js processes on these dynamically resolved paths constitutes dynamic execution of local code.
[INDIRECT_PROMPT_INJECTION]: The skill ingests content from external markdown files (docs/unwind/questions/*.md) designed to be modified by users or domain experts. The answers are then propagated back into the system's primary documentation layers (docs/unwind/layers/**).
Ingestion points: Processes docs/unwind/questions/*.md (expert answers) and docs/unwind/layers/** (layer documentation).
Boundary markers: Employs specific HTML comment markers () to delimit machine-readable metadata within natural language text.
Capability inventory: Utilizes Bash (git, node, pnpm), Write, Edit, and Task (subagent invocation) tools.
Sanitization: Performs basic validation of verdict checkboxes but does not sanitize the natural language 'Notes' or 'corrected behavior' fields which are written back to the documentation, creating a potential vector for instructions to reach downstream code generation tools.

uw-grill