AgentSkillsCN

deep-research

利用Gemini Deep Research Agent（Interactions API）开展自主的多步骤研究。支持网页搜索、文件/目录上下文，以及稳健的流式传输。

SKILL.md
--- frontmatter
name: deep-research
description: "Perform autonomous, multi-step research using the Gemini Deep Research Agent (Interactions API). Supports web search, file/directory context, and resilient streaming."

Deep Research

Use this skill to conduct deep, autonomous research tasks that require planning, searching, reading, and synthesizing information.

This skill uses the Gemini Deep Research Agent (deep-research-pro-preview-12-2025) via the Interactions API.

Prerequisites

  • •GOOGLE_API_KEY: Required for accessing the Interactions API (currently in Preview via AI Studio).
  • •google-genai SDK v0.3.0+

Usage

Basic Research

Start a research task and stream the results to the console.

bash
uv run skills/deep-research/scripts/research.py "Research the history of RISC-V architecture."

Research with Context (Files or Directories)

Provide local files (PDFs, text) or entire directories for the agent to read and incorporate into its research.

Direct Upload (Best for specific docs):

bash
# Single file
uv run skills/deep-research/scripts/research.py "Analyze this report" --file report.pdf

# Entire directory
uv run skills/deep-research/scripts/research.py "Summarize these meeting notes" --file ./notes/

File Search Store (Best for large corpora): Use --use-file-store to index files into a searchable store (RAG) instead of uploading them into the context window.

bash
uv run skills/deep-research/scripts/research.py "Find trends in these 1000 PDFs" --file ./large_corpus/ --use-file-store

Saving the Report

Save the final Markdown report to a file.

bash
uv run skills/deep-research/scripts/research.py "Competitive landscape of EV batteries" --output report.md

Continuing Research (Follow-up)

Ask follow-up questions to an existing research session using the Interaction ID (displayed at the start of the previous run).

bash
uv run skills/deep-research/scripts/research.py "Elaborate on the second point about lithium supply." --follow-up "INTERACTION_ID_HERE"

References

How it Works

  1. •Planning: The agent breaks down your prompt into steps.
  2. •Execution: It autonomously searches the web and reads your provided files.
  3. •Resilience: The script automatically reconnects if the long-running stream drops.
  4. •Synthesis: It produces a comprehensive, cited report.