AgentSkillsCN

deep_research

借助 Gemini 的高级控制功能（日志概率、引用、长上下文），开展严谨的学术研究与文献分析。

SKILL.md
--- frontmatter
name: deep_research
description: Conducts rigorous academic research and document analysis using Gemini's advanced control features (Logprobs, Citations, Long Context).
usage_trigger: Use when the user asks for "research", "paper analysis", "fact checking", "academic review", or "uncertainty analysis".

🔬 Deep Research Skill

Capabilities

This skill leverages specific Gemini Research features to provide higher-fidelity analysis than standard chat.

  1. •

    Uncertainty Quantification (Logprobs):

    • •When analyzing complex topics, this skill checks the logprobs of the generated tokens.
    • •It highlights statements where the model's confidence is low, preventing hallucinations.
  2. •

    Rigorous Citations (CitationMetadata):

    • •Enforces strict citation extraction using groundingMetadata.
  3. •

    Long-Context Synthesis:

    • •Optimized for ingesting full PDF papers or large text dumps (up to 2M tokens).

Instructions

For Paper Analysis

  1. •Ingest: Read the full content of the file.
  2. •Analyze: Use the configured Pro model (currently gemini-3-pro-preview) with thinking_level="HIGH".
  3. •Output: structured report with:
    • •Core Thesis: What is the paper claiming?
    • •Methodology Evaluation: Is it sound?
    • •Confidence Score: Based on Logprobs (simulated or actual if access permits).
    • •Novelty Assessment: How does this comparing to existing literature?

For Fact Checking

  1. •Activate google_search tool.
  2. •Cross-reference claims.
  3. •Explicitly list CitationMetadata sources.

Python Helper (Concept)

Use visions/core/visions_backend.py to access the raw client if you need to inspect response.candidates[0].logprobs.