The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages shell commands to perform repository analysis. The audit-changed-skills command uses git diff to dynamically determine which skills require auditing based on branch changes. Additionally, the skill's instructions recommend the use of python3 for compatibility checks and pnpm exec playwright for validating web-based skill outputs.
[EXTERNAL_DOWNLOADS]: The skill's documentation (README.md) fetches a static SVG asset from a public GitHub repository belonging to the author (github.com/sorawit-w/agent-skills). This is a trusted developer resource used for informational purposes.
[DATA_EXFILTRATION]: To perform its primary function, the skill accesses and reads the full content of SKILL.md files and all referenced files within the user's workspace. This broad file access is necessary for the audit but represents a significant data exposure surface.
[PROMPT_INJECTION]: As a tool designed to process and act upon instructions from third-party skills, the evaluator has an indirect prompt injection surface.
Ingestion points: Phase 1 involves reading the target skill's body and all referenced documentation or scripts.
Boundary markers: The skill uses a split-role harness (described in references/executor-brief.md and references/grader-brief.md) that employs fresh contexts and 'stated-intent' tool calling. This design ensures that the agent role-plays the audited skill's behavior without actually executing potentially malicious commands in the live environment.
Capability inventory: The skill is capable of spawning sub-agents, performing comprehensive workspace file reads, and executing shell commands via the git CLI.
Sanitization: The skill relies on 'stated-intent' suffixes for tool names and execution/grading role separation rather than explicit text sanitization or filtering.

skill-evaluator