The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as a meta-generator that ingests user domain requirements to architect and write new agent definitions and skills, creating a potential vector for malicious data to influence the generated components.\n
Ingestion points: The skill reads user requests to define the project domain and also audits existing configuration files in .claude/agents/, .claude/skills/, and CLAUDE.md to guide its generation process.\n
Boundary markers: The instructions and provided templates lack explicit requirements for using boundary markers, XML tags, or 'ignore nested instructions' prompts when interpolating user data into the generated skill descriptions or agent roles.\n
Capability inventory: The skill utilizes powerful tools including file writing (Write), team creation (TeamCreate), and task management (TaskCreate), all of which are orchestrated based on the processed input.\n
Sanitization: The skill does not define specific validation, sanitization, or escaping logic to ensure that user-provided domain details do not contain instructions that could hijack the generated agents' behavior.

harness-zh