The Agent Skills Directory

[DATA_EXFILTRATION]: The skill implements a security boundary for sensitive data (names, deals, private reasoning) by explicitly instructing the agent to acknowledge the data while keeping the actual substance within the chat context, ensuring it is never committed to a file.
[COMMAND_EXECUTION]: Provides a local helper script assets/input-alert.sh intended for human execution in a terminal. The script uses standard local utilities for ASCII art, terminal alerts, and text-to-speech (e.g., say, espeak). It does not perform network operations or access sensitive system paths.
[PROMPT_INJECTION]: The skill processes untrusted user-provided content, which represents an indirect prompt injection surface when the agent is asked to update persistent project instructions.
Ingestion points: Processes raw user payloads as described in SKILL.md.
Boundary markers: No specific delimiters or boundary instructions are defined for the input payload block.
Capability inventory: Authorized to modify the working tree, update backlog files, and propose edits to agent instruction files (e.g., AGENTS.md or CLAUDE.md).
Sanitization: Mitigation is provided through a mandatory human-in-the-loop workflow where the agent must 'propose' durable changes for user approval before they are committed.
[SAFE]: No malicious patterns such as obfuscation, credential harvesting, or unauthorized remote code execution were detected. The skill's behavior aligns with its stated purpose of data organization.

input