The Agent Skills Directory

[COMMAND_EXECUTION]: The file render-graphs.js utilizes the execSync function to invoke the system dot command (Graphviz). The script extracts Graphviz code from markdown files and pipes it directly into the binary, allowing for local command execution and potentially exposing the environment to vulnerabilities within the Graphviz software.
[PROMPT_INJECTION]: The skill uses 'persuasion principles' such as Authority and Commitment to steer agent behavior and override its internal reasoning processes. Directives such as 'YOU MUST', 'No exceptions', and 'Delete means delete' are explicitly designed to bypass agent rationalization and force compliance with specific process rituals.
[DATA_EXFILTRATION]: The skill's instructions and test scenarios frequently reference and encourage access to internal agent configuration paths, including ~/.claude/skills/ and ~/.agents/skills/. Accessing these directories exposes the agent's installed extensions and internal organizational structure.
[EXTERNAL_DOWNLOADS]: The anthropic-best-practices.md file provides instructions for downloading and installing external third-party libraries, such as pdfplumber and pypdf, using the pip install command.

writing-skills