The Agent Skills Directory

[COMMAND_EXECUTION]: The skill modifies the ".claude/settings.json" file to update the "permissions.allow" section. This action alters the core security policy of the environment by explicitly granting tool access to newly created agents.
[COMMAND_EXECUTION]: It utilizes the "/schedule" tool to create persistent, automated tasks that execute code in the background on a recurring basis.
[REMOTE_CODE_EXECUTION]: The skill dynamically generates and saves new instruction files (subagents) to ".claude/agents/" and then facilitates their execution, effectively creating and running new code at runtime.
[DATA_EXFILTRATION]: The skill systematically reads sensitive configuration and agent files, including ".claude/settings.json", ".claude/agents/*-farmer.md", and "context/watchlists.md", which contain metadata about the environment and other agent tasks.
[PROMPT_INJECTION]: The skill creates subagents that ingest data from external sources (such as Slack, calendars, and URLs) via MCP tools. The generated subagent instructions lack boundary markers and sanitization protocols, creating a significant surface for indirect prompt injection from malicious external data sources.

create-farmer