The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes and explicitly teaches the use of persuasion principles such as "Authority", "Commitment", and "Scarcity" to force agent compliance and override standard reasoning. It specifically instructs the use of preambles like "IMPORTANT: This is a real scenario. Choose and act" to bypass agent refusals or safety guardrails by framing instructions as reality rather than a prompt. The skill references research on persuading AI to comply with "objectionable requests" as a methodology for designing instructions.
[DYNAMIC_EXECUTION]: The render-graphs.js script uses child_process.execSync to execute the system dot command (Graphviz). It populates the input to this command with content extracted from SKILL.md files using regular expressions, which can lead to unexpected behavior if the dot binary is not configured securely.
[COMMAND_EXECUTION]: The tool relies on shell execution of the dot binary via execSync. While the command string itself is static, the input content is derived from untrusted Markdown blocks.
[INDIRECT_PROMPT_INJECTION]: The skill has a clear attack surface in render-graphs.js, which ingests untrusted Markdown data from the file system and processes it via the dot capability without sanitization or explicit boundary markers in the prompt logic.
Ingestion points: render-graphs.js reads SKILL.md content from local directories provided as command-line arguments.
Boundary markers: None found in the processing script or instructions to separate instructions from data.
Capability inventory: File system read/write via fs, command execution via execSync.
Sanitization: Content is extracted via regex but passed directly to the dot subprocess via stdin without validation of the Graphviz commands themselves.

writing-skills