The Agent Skills Directory

Prompt Injection (LOW): The documentation files (CLAUDE_MD_TESTING.md and persuasion-principles.md) explicitly use and teach authoritative override patterns.
Evidence: Phrases like 'THIS IS EXTREMELY IMPORTANT', 'YOU MUST', 'No exceptions', and 'If you didn't use it, you failed' are designed to override agent reasoning.
Context: These are presented as test variants and psychological principles for 'sticky' skill design rather than a malicious attack on the host agent.
Command Execution (LOW): The script render-graphs.js executes the system command dot (Graphviz).
Evidence: execSync('dot -Tsvg', { input: dotContent, ... }) is used to generate SVG files from markdown content.
Risk: The command is executed locally. The script mitigates injection risks by passing the diagram content via stdin rather than shell arguments and sanitizing output filenames using a \w+ regex match.
Indirect Prompt Injection (LOW): The skill establishes an ingestion surface for untrusted data and identifies subsequent capabilities.
Ingestion points: The skill is designed to process external SKILL.md files and browse the ~/.claude/skills/ directory.
Boundary markers: None identified; the skill encourages reading external files 'completely' without explicit delimiters.
Capability inventory: The skill has the ability to execute shell commands (via render-graphs.js) and write files to the local disk.
Sanitization: Filenames derived from external content are restricted to alphanumeric characters via regex.

writing-skills