The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from external files including .claude/skills/*/SKILL.md, .claude/agents/*.md, and catalog.yaml. This content is interpolated into the agent's context during auditing tasks. There are no instructions for the agent to use boundary markers or sanitization when processing this potentially untrusted data. Given the skill's access to powerful tools like Bash, Write, and Agent (sub-agents), a malicious instruction in a file being analyzed could trigger unauthorized actions.
Ingestion points: .claude/skills/*/SKILL.md (Phases 2A, 2B, 2D), .claude/agents/*.md (Phase 2B, 2D, 2E), catalog.yaml (Phase 2E), and unity-golden-tasks.md (Phase 2C).
Boundary markers: Absent.
Capability inventory: Read, Glob, Grep, Write, Bash, and Agent tools.
Sanitization: No sanitization or filtering of external instructions mentioned.
[COMMAND_EXECUTION]: The unity-eval mode and environment setup steps use the Bash tool for running unity, git, and installation commands. These commands rely on parameters like fixture paths and task IDs which could be manipulated.
[EXTERNAL_DOWNLOADS]: The skill instructions include a requirement to install Unity CLI if missing, implying the download of external components at runtime via the Bash tool.
[DYNAMIC_EXECUTION]: The unity-eval mode (Phase 2C) triggers the compilation and execution of code within a Unity project fixture to collect behavioral evidence, which involves running code derived from external project files.

skill-test