The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates indirect prompt injection by design through its UI feedback mechanism.\n
Ingestion points: The skill reads untrusted human feedback (comments) from a local server endpoint at http://localhost:4747/pending.\n
Boundary markers: Content is injected into the agent's context using text markers like === AGENTATION: {c} UI annotations ===, but it lacks explicit instructions for the agent to disregard instructions within those annotations.\n
Capability inventory: The skill is permitted to use Bash, Write, Read, and Grep, providing a significant impact surface if an injection succeeds.\n
Sanitization: The Python snippets provided in the SKILL.md hooks directly print the comment field into the agent output without any sanitization or escaping.\n- [COMMAND_EXECUTION]: The skill relies on shell command execution for its core lifecycle and platform integration.\n
Platform hooks for Claude Code and Gemini CLI use curl piped to python3 -c to process data from the local server.\n
The setup-agentation-mcp.sh script modifies agent-specific configuration files (e.g., ~/.claude/settings.json, ~/.gemini/settings.json) to register hooks and MCP servers.\n- [EXTERNAL_DOWNLOADS]: The skill uses package managers and remote installation tools to set up its environment.\n
It uses npx to dynamically download and execute the agentation-mcp package from the NPM registry.\n
It suggests installing the skill via npx skills add benjitaylor/agentation, which fetches code from a remote GitHub repository.

agentation