The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The browser-cdp module provides explicit instructions for the AI agent to extract sensitive session data, including browser cookies and authentication tokens, using the agent-browser tool via the Chrome DevTools Protocol (e.g., agent-browser --cdp 9222 eval 'localStorage.getItem("token") || document.cookie'). This represents a significant credential theft risk if the agent is misused.
[COMMAND_EXECUTION]: The skill relies extensively on running local JavaScript (node) and shell (bash) scripts provided within the skill package to manage project workflows, scrape data, and perform system tasks like killing browser processes (e.g., pkill -9 -x 'Google Chrome').
[COMMAND_EXECUTION]: The story-setup module performs deep modifications to the user's project environment by altering CLAUDE.md, updating .claude/settings.local.json, and installing git hooks (.git/hooks/pre-commit). This persistence-like behavior allows the skill to influence future agent interactions within the project.
[PROMPT_INJECTION]: There is a high risk of indirect prompt injection as the skill is designed to scrape and analyze data from multiple external web novel platforms (Qidian, Fanqie, Jjwxc, Qimao, Ciweimao). The ingested untrusted content is then used to generate project files, summaries, and writing suggestions.
[EXTERNAL_DOWNLOADS]: The skill fetches configuration and data from an external GitHub repository (github.com/worldwonderer/oh-story-claudecode) and interacts with OpenAI-compatible APIs, requiring the presence of sensitive API keys in the environment.

novel-assistant