The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches official Claude Code documentation, schema references, and plugin guidelines from code.claude.com to ensure created plugins comply with platform requirements.
[COMMAND_EXECUTION]: Utilizes local development tools including git for version control, jq for JSON manipulation, and grep for searching, as well as executing local Python scripts via uv for scaffolding.
[REMOTE_CODE_EXECUTION]: Leverages uvx to execute the skilllint package from a remote registry to perform automated validation of skill metadata and structure.
[DATA_EXFILTRATION]: Accesses the local plugins/ directory and the user's global skill repository at ~/.claude/skills/ to research existing implementations and ensure architectural consistency.
[PROMPT_INJECTION]: Vulnerable to indirect prompt injection because the skill ingests and analyzes third-party plugin code and external documentation.
Ingestion points: Content is read from local plugin files, the ~/.claude/skills/ directory, and external documentation URLs.
Boundary markers: There are no explicit delimiters or instructions provided to sub-agents to ignore potentially malicious instructions embedded in the ingested data.
Capability inventory: The orchestration involves sub-agents (general-purpose, Plan, plugin-assessor) that have permissions to write to the file system and execute shell commands.
Sanitization: The skill does not perform sanitization or validation of the logic contained within ingested files before they are processed by LLM sub-agents.

plugin-creator