The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied content (seed tasks or existing CLAUDE.md files) to refactor them into a new structured prompt. While the skill applies a 'Lint' process to remove emotional anti-patterns, it does not explicitly filter for malicious instructions (e.g., jailbreaks or system prompt extraction) that might be embedded in the source prompt being refactored. If a user provides a malicious prompt for improvement, the skill will attempt to preserve its core requirements in the output artifact.
Ingestion points: User input (seed tasks, pasted prompts, or files) identified in the 'Inventory the input' section of SKILL.md.
Boundary markers: The output is structured using markdown headers (e.g., ## Task, ## Context) as defined in assets/template.md, but it does not utilize specific isolation tokens (like XML tags or delimiters) to wrap the refactored content for the downstream agent.
Capability inventory: The skill possesses file-writing capabilities (write the file directly) which are triggered when a user supplies a destination path.
Sanitization: The skill performs a phrasing lint against references/eip-anti-patterns.md to ensure the final prompt uses collaborative rather than authoritarian language, but this is a stylistic transformation rather than a security filter.

prompt