AgentSkillsCN

research

根据给定的研究主题,查找相关科学论文,分析其研究方法,并提供结构化的摘要,概述整体研究方法。

SKILL.md
--- frontmatter
name: research
description: Find relevant scientific papers for a given research prompt, analyze their approaches, and provide structured summaries with a general methodology overview.
user-invocable: true
allowed-tools: Read, Write, Glob, Grep, WebSearch, WebFetch, AskUserQuestion

Research Skill

Find, analyze, and summarize scientific papers relevant to a research prompt. Produces structured per-paper summaries and a general methodology overview.

Usage

  • /research - Start interactive research session (prompts for topic)
  • /research "cross-lingual document retrieval using embedding models" - Research a specific topic
  • "Find papers on low-resource NER"
  • "Research recent approaches to knowledge distillation for LLMs"

Workflow

Step 1: Extract Topics

From the user's prompt, identify 3-6 specific research topics or keywords that capture the core interests. Present these as interactive checkboxes using AskUserQuestion so the user can confirm, deselect, or add topics.

Example:

code
From your prompt, I identified these topics. Which ones should I search for?

[ ] Cross-lingual retrieval
[ ] Dense passage embeddings
[ ] Multilingual language models
[ ] Document ranking
[ ] Zero-shot transfer

Wait for user confirmation before proceeding. If the user adds custom topics, incorporate them.

Step 2: Search for Papers

Search for papers matching the confirmed topics. Use WebSearch with academic queries.

Search strategy:

  • Combine confirmed topics into 2-4 targeted queries
  • Prefer peer-reviewed venues: ACL, EMNLP, NAACL, NeurIPS, ICML, ICLR, AAAI, CVPR, ICCV, ECCV, IEEE, Elsevier, Springer
  • Include venue/journal names in queries to bias toward peer-reviewed work
  • Add "paper" or "proceedings" to queries
  • Search for recent work (last 2-3 years) and seminal/highly-cited work
  • Aim to find 5-10 relevant papers

Query patterns:

  • "topic1 topic2" site:aclanthology.org OR site:arxiv.org OR site:openreview.net
  • "topic1" "topic2" conference paper 2024 2025
  • "topic1" survey OR review journal

For each paper found, fetch its abstract and metadata using WebFetch when possible.

Step 3: Assess Relevancy

For each paper, assign a relevancy score on a 1-5 scale based on how well it matches the confirmed topics:

ScoreMeaning
5Directly addresses the core prompt; covers multiple confirmed topics
4Highly relevant; addresses most confirmed topics
3Moderately relevant; covers some topics or a closely related area
2Tangentially relevant; shares methodology or domain but different focus
1Loosely related; only peripherally connected to the prompt

Sort papers by relevancy score (highest first). Only include papers scoring 3 or above in the final output. Mention excluded papers briefly if they were borderline.

Step 4: Summarize Each Paper

For each paper (sorted by relevancy), provide a structured summary using this exact format:

markdown
## Paper Title (Year)
**Authors:** Author list
**Venue:** Conference/Journal name
**Relevancy:** X/5

### Motivation
- Bullet point 1
- Bullet point 2

### Problem to Solve
- Bullet point 1
- Bullet point 2

### Methodology
- Bullet point 1
- Bullet point 2
- Bullet point 3

### Datasets & Metrics
- Bullet point 1 (dataset details)
- Bullet point 2 (metrics used)

### Results & Insights
- Bullet point 1
- Bullet point 2
- Bullet point 3

Summary rules:

  • Keep each section to 2-4 bullet points
  • Be concise: one idea per bullet, no filler
  • Use concrete numbers for results when available
  • For methodology, focus on what makes their approach novel
  • For results, highlight comparisons to baselines

Step 5: General Methodology Summary

After all individual summaries, provide a General Summary section that synthesizes the findings across all papers. This should NOT reference specific paper names or methods. Instead, describe general patterns and effective strategies.

markdown
# General Summary

## Effective Approaches
- General description of what types of changes/methods work well
- Patterns observed across successful approaches

## Key Factors
- What aspects of the problem seem most important to address
- Common components in high-performing solutions

## Open Challenges
- What problems remain unsolved
- Where current approaches fall short

## Recommendations
- Based on the surveyed work, what directions seem most promising
- What combinations of techniques appear effective

General summary rules:

  • Abstract away from specific paper names: say "approaches that incorporate X" not "Method Y from Paper Z"
  • Focus on transferable insights: "adapting pre-trained representations to the target domain improves..." not "BERT-based fine-tuning gave 3% improvement"
  • Identify convergent trends across papers
  • Highlight disagreements or contrasting findings
  • Keep each section to 3-5 bullet points

Output Format

The complete output should follow this structure:

markdown
# Research Summary: [Topic from prompt]

**Topics searched:** topic1, topic2, topic3
**Papers found:** N total, M included (relevancy >= 3)
**Date of search:** YYYY-MM-DD

---

[Individual paper summaries, sorted by relevancy score]

---

# General Summary

[Synthesized methodology overview]

Tips

  • If a paper's full text is not accessible, summarize based on the abstract and any available information
  • When results are ambiguous or the abstract lacks detail, note this explicitly
  • Prefer papers from top-tier venues when relevancy scores are equal
  • Include survey/review papers if found, as they provide broader context
  • If fewer than 3 papers are found, suggest broadening or adjusting the topics