The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as a gate for persisting user-provided information into permanent instruction files such as CLAUDE.md, AGENTS.md, and MEMORY.md. These files are loaded into the agent's context in subsequent sessions.
Ingestion points: Data enters the agent's context when a user provides instructions or facts to be "remembered" or added to the project documentation.
Boundary markers: The skill does not define specific delimiters or warnings to separate user-provided data from the agent's core instructions, potentially allowing injected content to influence future behavior.
Capability inventory: The agent is authorized to use Edit, Write, and shell commands like echo, sed, and tee to persist this data.
Sanitization: The skill mitigates risk by requiring the agent to condense input into specific architectural patterns (e.g., "X not Y because Z") and routing it to appropriate files, which provides a layer of manual logic filtering.
[COMMAND_EXECUTION]: The instructions guide the agent to use shell-based write commands (echo >>, sed -i, tee -a, heredocs) to modify system-level and project-level instruction files. The skill acknowledges that these methods may bypass certain platform hooks that monitor standard file-editing tools.
[PERSISTENCE]: The skill utilizes a coordination mechanism involving a "per-project flag" stored in the system temporary directory (tmpdir). This mechanism allows the agent to maintain state and suppress reminders across different tool invocations and skill transitions.

instruction-guardian