deep-research_skill

Overview

This skill performs autonomous, multi-step research using the Gemini Deep Research Agent via the Interactions API. It plans tasks, executes web searches, ingests local files or directories, and synthesizes a cited Markdown report. It also supports resilient streaming and follow-up interactions for iterative research.

How this skill works

The agent first decomposes your prompt into a step-by-step plan. It then executes web searches and reads any provided files or indexed file stores, streaming progress as it works. If the streaming connection drops, the tool automatically reconnects and continues. The final output is a synthesized, cited report you can save to disk or continue with follow-up queries.

When to use it

• When you need a comprehensive research report that combines web search and local documents.
• When analyzing large document collections (PDFs, text) for trends or summaries.
• When you want an automated multi-step investigation with citations and structure.
• When you need resilient long-running research that can reconnect and continue.
• When you plan to iterate using follow-up queries tied to a prior interaction session.

Best practices

• Provide clear, specific research questions and scope to focus the agent’s planning phase.
• Upload small, representative documents directly; use the file-store index for large corpora.
• Include relevant directories rather than individual files when multiple documents share context.
• Save and reuse the Interaction ID to follow up or refine prior research results.
• Request the final output in Markdown if you intend to edit or publish the report.

Example use cases

• Create a cited historical overview and timeline for a technology like RISC-V.
• Summarize and extract trends across hundreds of industry PDFs using a file-store index.
• Produce a competitive landscape report and save it as a Markdown deliverable.
• Ingest internal meeting notes and generate an executive summary with action items.
• Run iterative deep dives: start broad, then ask follow-ups referencing the Interaction ID.