The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines an Autopilot mode via the --auto flag that explicitly directs the agent to "Skip ALL user confirmations" and execute a multi-phase research pipeline end-to-end. This instruction overrides standard safety practices that require human verification before executing shell commands and spawning autonomous tool agents.
[COMMAND_EXECUTION]: The orchestrator uses the Bash tool to perform system-level tasks including session initialization, file creation, and the execution of local shell scripts (gemini-research.sh, browser-extract.sh). These scripts invoke external CLI tools like chromux and gemini to perform research operations.
[EXTERNAL_DOWNLOADS]: The skill utilizes the npx tool to dynamically execute @team-attention/chromux from the NPM registry. It also incorporates the gemini CLI, a tool provided by a well-known service (the google-gemini organization), for independent research tasks.
[PROMPT_INJECTION]: The skill presents an Indirect Prompt Injection surface by retrieving and processing content from untrusted external web sources.
Ingestion points: Untrusted content enters the agent's context through WebSearch, WebFetch, and browser-based data extraction agents (SKILL.md).
Boundary markers: The instructions do not include boundary markers or delimiters to isolate retrieved web data from the agent's own instructional context.
Capability inventory: The skill is capable of executing arbitrary shell commands via Bash and spawning sub-agents with tool-use permissions (SKILL.md).
Sanitization: The workflow does not include any steps to sanitize or validate external content before it is processed by the agent or used to generate the final report.

deep-research