The Agent Skills Directory

[SAFE]: The skill was analyzed for malicious patterns including prompt injection, data exfiltration, and obfuscation. No malicious instructions or hidden payloads were found. The technical content accurately reflects psychological and AI research principles.
[EXTERNAL_DOWNLOADS]: The skill includes numerous external references in its YAML metadata (grounding.truth_sources). These links target trusted organizations (Anthropic, OpenAI, Google) and well-known academic/technical services (arXiv, DOI, Crossref). These resources are used for theory grounding and do not involve remote code execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze and process untrusted external content, such as skill bodies or prompts written by third parties. While this presents a potential attack surface, the skill includes explicit diagnostic kernels and protocols to detect 'cognitive overload attacks'—where malicious actors use excessive context to confuse a model's safety filters.
Ingestion points: Analyzes SKILL.md files, prompt strings, and documentation inputs.
Boundary markers: Explicit audit steps provided in the 'Load Audit Kernel' and 'Diagnostic Protocol'.
Capability inventory: Limited to Read and Grep tools as defined in the frontmatter.
Sanitization: The skill emphasizes structural classification (Intrinsic vs. Extraneous) as a method to filter noise from substance.

cognitive-load-theory