The Agent Skills Directory

[SAFE]: The skill is composed of markdown instructions and a large corpus of text-based research data (transcripts, summaries, and bibliographies) used to inform the agent's reasoning. No executable scripts, binaries, or dangerous system commands were found.
[EXTERNAL_DOWNLOADS]: All URLs and external references discovered in the reference files and source data point to highly reputable and established educational or social platforms, including stanford.edu, github.com, youtube.com, wikipedia.org, and arxiv.org. These are used strictly for attribution and further learning.
[PROMPT_INJECTION]: The instructions do not contain behavior-override patterns. In fact, the skill instructs the agent to adopt a skeptical and empirical stance toward AI capabilities, highlighting 'Jagged Intelligence' and requiring 'Human Verification' for generated outputs.
[DATA_EXFILTRATION]: No instructions or patterns related to harvesting credentials or sensitive local files were detected. The skill's primary function is processing public technical knowledge.
[COMMAND_EXECUTION]: There is no use of dynamic context injection (exclamation-backtick syntax) or instructions to perform unauthorized shell operations. The mention of tools like 'Code Interpreter' is in the context of recommended standard LLM features for math and technical verification.

andrej-karpathy