The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references and requires the installation of the gemini CLI from Google's official GitHub repository (github.com/google/gemini-cli). As this is a well-known service provided by a trusted organization, this dependency is considered safe and used as intended for the skill's 'Lite' mode.
[COMMAND_EXECUTION]: The skill executes external commands to perform its primary functions:
In scripts/lite_research.sh, it invokes the gemini CLI tool with user-provided topics and outlines to generate research drafts.
In SKILL.md, it describes the use of python and bash to run the provided research scripts.
These executions are part of the intended functionality and use standard shell quoting for variables like $TOPIC and $MODEL.
[PROMPT_INJECTION]: The skill exhibits an attack surface for Indirect Prompt Injection (Category 8) due to its core functionality of processing external data:
Ingestion points: Untrusted data enters the agent context through the data_dir (scanned recursively in deep_research.py), attachments (PDFs, images, videos), and the outline.md file.
Boundary markers: The prompt construction in build_research_prompt and lite_research.sh uses structural labels (e.g., 'Research topic:', 'Research提纲') to delimit content, but lacks explicit instructions to the AI to ignore instructions embedded within the processed files.
Capability inventory: The Gemini Deep Research agent has access to google_search, url_context, and file_search (via Google's infrastructure). It can also generate reports and store sessions.
Sanitization: No sanitization or filtering is performed on the content of files uploaded to the FileSearchStore or passed as attachments.

gemini-deepresearch