The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to install external CLI tools (claudish, claudeup) globally using npm install -g. These tools are utilized for managing MCP servers and performing external visual analysis calls.
[DYNAMIC_EXECUTION]: The skill leverages the mcp__chrome-devtools__evaluate_script tool, which allows the execution of arbitrary JavaScript within the context of the browser page being debugged. Additionally, it uses npx claudish to dynamically generate and execute prompts against external vision models.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it ingests untrusted data from the DOM, console messages, and network requests of external web pages. This data is then interpolated into prompts for visual analysis models, which could be manipulated by malicious website content to produce deceptive debugging reports.
Ingestion points: Page snapshots, screenshots, console logs, and network request details via Chrome DevTools MCP.
Capabilities: The agent can execute shell commands (Bash) to run analysis tools and write to local configuration files.
Boundary markers: Limited use of delimiters (e.g., IMAGE: path) in generated prompts.
Sanitization: No explicit sanitization of console output or page text before processing.
[PERSISTENCE]: The skill performs persistent storage of user preferences by modifying .claude/settings.json and reading/writing to session metadata in ai-docs/sessions/.
[DATA_EXFILTRATION]: Captured screenshots of implementation work (potentially including local development environments and sensitive UI data) are sent to external APIs (OpenRouter) and model providers (Qwen, Google, OpenAI) for analysis.

browser-debugger