AI文献报告生成器：自动搜索分析学术论文，生成结构化文献综述

Asta Literature Reports by allenai/asta-plugins

3 GitHub Stars

GitHub

安装命令

npx skills add https://github.com/allenai/asta-plugins --skill 'Asta Literature Reports'

AI/机器学习内容创作科研工具

🇨🇳中文介绍

文献报告生成

通过搜索学术文献、分析论文并将发现综合成结构良好的 Markdown 文档，生成全面的文献报告。

何时使用此技能

用户要求进行文献综述或全面报告
任务需要将多篇论文综合成一个连贯的叙述
用户想要一份书面交付成果（不仅仅是论文列表）
主题需要深入探索，并附带引用和证据
用户已有现有的论文查找结果，并希望从中生成一份全面的报告

工作流程

1. 检查现有结果

如果用户提供了现有的论文查找结果文件：

请用户确认文件路径
使用 Read 工具验证文件是否存在且为有效的 JSON
直接跳到步骤 3 处理这些结果

如果没有现有结果文件：

继续执行下面的新搜索

2. 启动后台搜索（如果需要）

首先在后台启动论文查找器（30-60秒）：

# 使用默认位置 .asta/literature/find/YYYY-MM-DD-HH-MM-SS-{query-slug}.json
Bash(command="asta literature find 'query' --timeout 300", run_in_background=true)

# 或指定自定义输出位置
Bash(command="asta literature find 'query' --timeout 300 -o custom-results.json", run_in_background=true)

在论文查找器运行的同时，根据需要启动其他有针对性的 CLI 搜索。

3. 立即创建报告文件

不要等待搜索完成。尽早创建报告文件：

默认文件路径：.asta/literature/report/YYYY-MM-DD-topic-slug.md（在当前工作目录中）

广告位招租

在这里展示您的产品或服务

触达数万 AI 开发者，精准高效

联系我们

4. 处理论文查找器结果

如果使用现有结果文件：

使用用户提供的文件路径
使用 Read 工具读取文件，以了解查询和论文信息

如果启动了新搜索：

当收到 <task-notification> 时，使用 TaskOutput 获取结果：

TaskOutput(task_id="<id>")

结果文件将位于 .asta/literature/find/（检查 TaskOutput 以获取确切的文件名），或者位于您使用 -o 指定的路径

使用 jq 浏览结果（按相关性排序）：

# 将 RESULTS_FILE 替换为 TaskOutput 中的实际路径或您的自定义路径
jq '[.results | sort_by(-.relevanceScore) | .[0:10][] | {title, year, venue, corpusId, citationCount, score: .relevanceScore, authors: (.authors | if length <= 4 then [.[].name] | join(", ") else ([.[0:3][].name] | join(", ")) + ", +" + (length - 4 | tostring) + ", " + .[-1].name end), summary: .relevanceJudgement.relevanceSummary}]' RESULTS_FILE

重要提示：浏览所有高度相关的论文（不仅仅是前几篇）。提取：

相关性标准和证据
论文正文中的片段
引用上下文（其他人如何引用这项工作）

5. 深入研究关键论文

对于每篇重要的论文，提取详细的证据：

相关性证据：

jq '.results[] | select(.corpusId == CORPUS_ID) | {title, criteria: [.relevanceJudgement.relevanceCriteriaJudgements[] | {concept: .name, relevance, evidence: [.relevantSnippets[].text]}]}' RESULTS_FILE

摘要和段落：

jq '.results[] | select(.corpusId == CORPUS_ID) | {title, abstract, snippets: [.snippets[]? | {section: .sectionTitle, text}]}' RESULTS_FILE

引用上下文（其他人如何描述这项工作）：

jq '[.results[] | select(.corpusId == CORPUS_ID) | .citationContexts[].text]' RESULTS_FILE

将 RESULTS_FILE 替换为实际的结果文件路径（位于 .asta/literature/find/、用户提供或来自 TaskOutput）。

6. 使用 CLI 工具进行针对性研究

使用 asta papers 进行快速、有针对性的搜索，以补充论文查找器的结果：

# 搜索论文
asta papers search "keyword" --year 2023- --limit 20 --fields title,abstract,year,authors

# 获取论文详情
asta papers get ARXIV:2005.14165 --fields title,abstract,authors,year

# 获取引用
asta papers citations CorpusId:12345 --limit 50

# 作者论文
asta papers author search "Author Name"
asta papers author papers <author_id> --limit 50

初始搜索使用较小的限制（10-20）
推荐字段：title,abstract,year,authors,venue,citationCount
使用 --format json 以管道传输到 jq，使用 --format text 进行快速查看

7. 逐步更新报告

在收集信息时：

立即将论文添加到参考文献部分
随着主题的出现，填写章节内容
随着理解的深入，更新执行摘要
使用正确的引用格式（见下文）

不要等到最后才写所有内容。 逐步编辑。

8. 完成前确保完整性

如果启动了后台搜索：所有搜索必须完成（检查是否有待处理的 <task-notification>）
所有章节必须填写完整
所有文内引用必须与参考文献匹配
执行摘要需反映关键发现
在底部添加生成日期

论文查找器结果使用 LiteratureSearchResult 格式（完整的结构文档请参见 查找文献 技能）。

报告生成的关键字段：

.query - 搜索查询
.results[] - 按相关性排序的论文数组
.results[].relevanceScore - 0-1 相关性分数
.results[].relevanceJudgement.relevanceSummary - AI 解释
.results[].relevanceJudgement.relevanceCriteriaJudgements[] - 每个概念的证据
.results[].snippets[] - 论文正文的文本摘录
.results[].citationContexts[] - 来自其他论文的引用上下文

使用带有链接定义的引用键来实现可点击的引用：

This was demonstrated by [[Yao2024]].

参考文献部分：

## 参考文献

- [[Yao2024]] Yao, S., et al. (2024). τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv.
- [[Dou2025]] Dou, Z., et al. (2025). Another Paper Title. Venue Name.

链接定义（在文件末尾）：

[Yao2024]: https://semanticscholar.org/p/270218537
[Dou2025]: https://semanticscholar.org/p/123456789

这使得 [[Yao2024]] 渲染为带有可见括号的 [Yao2024]，并可点击链接到 Semantic Scholar。

全面性：涵盖所有主要主题和关键论文
基于证据：用论文证据和引用来支持主张
综合性：不仅仅是列出论文——要连接不同论文的观点
结构良好：清晰的章节，逻辑流畅
引用规范：所有主张都能追溯到来源
内容最新：在相关时关注近期工作

🇺🇸English

Literature Report Generation

Generate comprehensive literature reports by searching academic literature, analyzing papers, and synthesizing findings into a well-structured markdown document.

When to Use This Skill

User asks for a literature review or comprehensive report
Task requires synthesizing multiple papers into a coherent narrative
User wants a written deliverable (not just a list of papers)
Topic requires deep exploration with citations and evidence
User has existing paper-finder results and wants a comprehensive report from them

Workflow

1. Check for Existing Results

If the user provides an existing paper-finder results file:

Ask them to confirm the file path
Use Read tool to verify the file exists and is valid JSON
Skip to step 3 to process those results directly

If no existing results file:

Proceed to launch a new search below

2. Launch Background Searches (if needed)

Start with the paper finder in background (30-60s):

# Uses default location .asta/literature/find/YYYY-MM-DD-HH-MM-SS-{query-slug}.json
Bash(command="asta literature find 'query' --timeout 300", run_in_background=true)

# Or specify a custom output location
Bash(command="asta literature find 'query' --timeout 300 -o custom-results.json", run_in_background=true)

Launch additional targeted CLI searches as needed while paper finder runs.

3. Create Report File Immediately

Don't wait for searches to complete. Create the report file early with:

Default file path: .asta/literature/report/YYYY-MM-DD-topic-slug.md (in current working directory)
Custom path: User can specify a different location if desired
Initial structure: Title, Executive Summary (placeholder), section headings, References

This keeps the user informed and avoids losing work.

4. Process Paper Finder Results

If using existing results file:

Use the file path provided by the user
Read it with the Read tool to understand the query and papers

If launched a new search:

When <task-notification> arrives, use TaskOutput to get the results:

TaskOutput(task_id="<id>")
The results file will be at .asta/literature/find/ (check TaskOutput for the exact filename), or at the path you specified with -o

Browse results with jq (sorted by relevance):

# Replace RESULTS_FILE with the actual path from TaskOutput or your custom path
jq '[.results | sort_by(-.relevanceScore) | .[0:10][] | {title, year, venue, corpusId, citationCount, score: .relevanceScore, authors: (.authors | if length <= 4 then [.[].name] | join(", ") else ([.[0:3][].name] | join(", ")) + ", +" + (length - 4 | tostring) + ", " + .[-1].name end), summary: .relevanceJudgement.relevanceSummary}]' RESULTS_FILE

Important : Go through ALL highly relevant papers (not just top few). Extract:

Relevance criteria and evidence
Snippets from paper body
Citation contexts (how others cite this work)

5. Deep Dive on Key Papers

For each important paper, extract detailed evidence:

Relevance evidence :

jq '.results[] | select(.corpusId == CORPUS_ID) | {title, criteria: [.relevanceJudgement.relevanceCriteriaJudgements[] | {concept: .name, relevance, evidence: [.relevantSnippets[].text]}]}' RESULTS_FILE

Abstract and passages :

jq '.results[] | select(.corpusId == CORPUS_ID) | {title, abstract, snippets: [.snippets[]? | {section: .sectionTitle, text}]}' RESULTS_FILE

Citation contexts (how others describe this work):

jq '[.results[] | select(.corpusId == CORPUS_ID) | .citationContexts[].text]' RESULTS_FILE

Replace RESULTS_FILE with the actual results file path (found in .asta/literature/find/, user-provided, or from TaskOutput).

6. Use CLI Tools for Targeted Research

Supplement paper finder with fast, targeted searches using asta papers:

# Search for papers
asta papers search "keyword" --year 2023- --limit 20 --fields title,abstract,year,authors

# Get paper details
asta papers get ARXIV:2005.14165 --fields title,abstract,authors,year

# Get citations
asta papers citations CorpusId:12345 --limit 50

# Author papers
asta papers author search "Author Name"
asta papers author papers <author_id> --limit 50

Tips:

Use smaller limits (10-20) for initial searches
Recommended fields: title,abstract,year,authors,venue,citationCount
Use --format json for piping to jq, --format text for quick viewing

7. Update Report Progressively

As you gather information:

Add papers to References section immediately
Fill in section content as themes emerge
Update Executive Summary as understanding develops
Use proper citation format (see below)

Do not wait until the end to write everything. Edit progressively.

8. Ensure Completeness Before Finishing

Before completing:

If launched background searches: All searches must finish (check for pending <task-notification>)
All sections must be filled in
All in-text citations must match references
Executive summary reflects key findings
Add generation date at bottom

JSON Structure

Paper finder results use the LiteratureSearchResult format (see Find Literature skill for complete structure documentation).

Key fields for report generation:

.query - the search query
.results[] - array of papers sorted by relevance
.results[].relevanceScore - 0-1 relevance score
.results[].relevanceJudgement.relevanceSummary - AI explanation
.results[].relevanceJudgement.relevanceCriteriaJudgements[] - per-concept evidence
.results[].snippets[] - text excerpts from paper body
.results[].citationContexts[] - citation contexts from other papers

Citation Format

Use citation keys with link definitions for clickable citations:

Inline citations :

This was demonstrated by [[Yao2024]].

References section :

## References

- [[Yao2024]] Yao, S., et al. (2024). τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv.
- [[Dou2025]] Dou, Z., et al. (2025). Another Paper Title. Venue Name.

Link definitions (at end of file):

[Yao2024]: https://semanticscholar.org/p/270218537
[Dou2025]: https://semanticscholar.org/p/123456789

This makes [[Yao2024]] render as [Yao2024] with visible brackets and clickable links to Semantic Scholar.

Quality Standards

Comprehensive : Cover all major themes and key papers
Evidence-based : Support claims with paper evidence and citations
Synthesized : Don't just list papers - connect ideas across papers
Well-structured : Clear sections, logical flow
Properly cited : All claims traced to sources
Up-to-date : Focus on recent work when relevant

Weekly Installs

–

Repository

allenai/asta-plugins

GitHub Stars

First Seen

–

Security Audits

Gen Agent Trust HubPass SocketPass SnykWarn

超能力技能使用指南：AI助手技能调用优先级与工作流程详解

45,100 周安装