The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script skills/render-html/scripts/render_html.py to process generated Markdown into HTML files.- [DATA_EXFILTRATION]: The instructions include a hardcoded local absolute path (/Users/yangruofeng/Desktop/aris_paper_discussion/aris_repo/docs/tutorials/attention_tutorial.md) as a style reference. This path is included in the instructions sent to an external model during the review process, exposing the local directory structure of the developer environment.- [PROMPT_INJECTION]: The skill ingests user-provided topics and author bylines which could contain malicious instructions. It mitigates indirect prompt injection risks by routing all generated content through a two-stage 'cross-model review' using an external high-reasoning model before any final files are produced.
Ingestion points: <topic> and --byline arguments in SKILL.md.
Boundary markers: Strict section skeletons and heading formats defined in the style guide.
Capability inventory: File writing, editing, and shell command execution (python3).
Sanitization: Mandatory verification rounds with an independent model to detect factual errors and style violations.- [SAFE]: Implements a strict security policy against personal information leaks, including a specific check in the review prompt for identifiers like server names, local paths, and affiliation details.- [SAFE]: Defaults to a non-autonomous mode (--commit false), ensuring that the agent does not push changes to a repository without manual user intervention.

interview-cheatsheet