The Agent Skills Directory

[COMMAND_EXECUTION]: The skill functions as a meta-orchestrator, executing various local and external CLI tools. It calls git for branch management, executes repository-specific linting and validation scripts (lint-skills.sh, validate-spec.sh), and invokes numerous AI harnesses (such as claude, codex, agy, and hermes) to perform peer reviews and behavioral testing.
[DYNAMIC_EXECUTION]: In references/harness-detection.md, the skill implements a three-step probe (PATH check, config hint, and smoke test) to detect available AI harnesses. It dynamically constructs shell commands from a predefined table and executes them to verify the functionality and model identity of these harnesses.
[INDIRECT_PROMPT_INJECTION]: Because the skill processes and analyzes the text of other AI skills (the 'candidates'), it has a significant surface for indirect prompt injection. The instructions include a mandatory defensive rule (Rule 14) stating that all candidate content must be treated as untrusted data, specifically forbidding the agent from following scoring instructions or quality signals embedded within the analyzed content. It further requires boundary markers and independent adjudication contexts to mitigate this risk.

skill-refiner