The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process user-provided novel text and persistent author preferences, which are subsequently used in subagent prompt templates. Ingestion points include local novel files (e.g., 正文.md) and author memory state files stored in the .story/ directory. The instruction templates for subagents, such as those found in references/agent-calls.md, do not include explicit security boundary markers or instructions to disregard potential commands embedded within the text content, making it susceptible to indirect prompt injection. Severity is assessed as low as this is a common attack surface for LLM-based content processors.\n- [COMMAND_EXECUTION]: The skill executes bundled local Node.js and Python scripts to perform automated text analysis and manage author style state. Specifically, scripts such as scripts/check-ai-patterns.js, scripts/check-degeneration.js, scripts/normalize-punctuation.js, and scripts/author_memory_commit.py are invoked during operation. These scripts are developed by the skill author and operate on local files within the project workspace, representing standard tool invocation within a controlled environment.\n- [SAFE]: The skill performs benign environment discovery by checking for the existence of platform-specific agent configuration directories (e.g., .claude/agents/, .opencode/agents/). This behavior is used to determine subagent compatibility and choose the appropriate invocation method for the current runtime environment, and it does not involve the harvesting or exfiltration of sensitive user credentials or data.

story-deslop