The Agent Skills Directory

PROMPT_INJECTION (HIGH): The files persuasion-principles.md and CLAUDE_MD_TESTING.md contain patterns designed to override agent behavior and bypass typical reasoning. Specifically, they advocate for the use of "Authority" and "Commitment" principles (e.g., "YOU MUST", "No exceptions", "If you didn't use it, you failed") to force compliance even under pressure or when the agent has a working alternative.
COMMAND_EXECUTION (MEDIUM): The script render-graphs.js uses child_process.execSync to run the system dot command. While it passes data via stdin, executing system binaries on data extracted from external files is a security risk if the binary has known vulnerabilities.
INDIRECT_PROMPT_INJECTION (HIGH):
Ingestion points: render-graphs.js reads content from SKILL.md files using the fs module.
Capability inventory: The script executes the dot binary via execSync to process the extracted content.
Boundary markers: None present. The script uses a simple regex to identify blocks and does not use any delimiters to separate instructions from data for the utility.
Sanitization: None. The script extracts raw strings and pipes them directly into a system subprocess.
METADATA_POISONING (MEDIUM): CLAUDE_MD_TESTING.md serves as a blueprint for modifying the agent's core operational instructions (CLAUDE.md) to ensure strict obedience to external skill files, potentially sidelining the agent's own safety or logic protocols.

writing-skills