The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides structured evaluation criteria and quality constraints. While it includes a 'NEVER Do When Evaluating' section, these are performance and quality instructions (e.g., 'NEVER give high scores just because it looks professional') rather than attempts to bypass AI safety guardrails or system instructions.
[DATA_EXFILTRATION]: No network operations, API calls, or hardcoded credentials were found. The skill does not access sensitive local file paths.
[OBFUSCATION]: The text was analyzed for Base64, zero-width characters, homoglyphs, and hidden text patterns. No obfuscation techniques were detected.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze other skill files. However, the skill-judge itself possesses no dangerous tools or capabilities (such as shell execution or network access) that could be exploited by malicious content within the audited files.
Ingestion points: Reads external SKILL.md files during audit tasks.
Boundary markers: None explicit in this instructional set, though it provides a scoring protocol.
Capability inventory: No tool usage, network operations, or file system write capabilities identified.
Sanitization: Not applicable as the skill is purely analytical and does not execute input.
[DYNAMIC_CONTEXT_INJECTION]: The file does not contain any dynamic shell execution patterns (!command).

skill-judge