shawncsx

AutoSCI — Data skill for Claude Code

Data community

AutoSCI — Distributed research-agent pipeline reproducing OpenAI's math-project workflow, retargeted at microbiome AI.

How to install AutoSCI

This entry records only its repository, not the path inside it, so there is no exact command to give. Open shawncsx/AutoSCI and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What AutoSCI does

AutoSCI — Distributed research-agent pipeline reproducing OpenAI's math-project workflow, retargeted at microbiome AI. Master / 10 sub-agents / 5 sources / 7-stage verification / Claude-Code-style rich CLI.

Alternatives in Data

  • Claude Deep Research Skill — Enterprise-grade deep research skill for Claude Code with 8-phase pipeline, source credibility scoring, and au 1k ★
  • Bug Bounty Master Workflow — Full pipeline: Recon -> Learn -> Hunt -> Validate -> Report 924 ★
  • Graph Engineering — Graph engineering for AI agents: the 9-stage knowledge-graph pipeline (translated from SEU's graduate course) 452 ★

README

AutoSCI — 分布式研究 Agent 流水线 / Distributed Research-Agent Pipeline

**复刻 OpenAI 的"数学论文工厂"流水线**,但跑在单机 Windows 上、面向 AI + 生信(微生物组表示学习)场景。

*A reverse-engineered clone of OpenAI's "math manuscript factory" — scaled down to a single Windows machine, retargeted at AI + bioinformatics (microbiome representation learning).*

AutoSCI 是一个 ~2,500 行的 Python 包。给定本地论文语料 + 一批种子 idea,它会**对每个 idea 串行调度 10 个专业子 agent**,产出一份"硕士级别"研究草稿(IMRaD 完整结构)+ 伪代码骨架 + 内部审计报告。整个设计是 OpenAI 数学项目工作流的可执行复刻 —— OpenAI 用这套工作流从 ~4,000 个候选问题里产出 719 篇论文:**Director → Research Team → keep / retry / drop**,每个子 agent 的职责被切得足够窄以保证稳定。

**核心思想**:OpenAI 的瓶颈从来不在模型,而在模型**周围**的工作流。AutoSCI 试图把这个工作流用 ~2,500 行 Python 写出来,让任何研究组都能在自己笔记本上批量产硕士级论文。


✨ 三大亮点

1️⃣ 调度官 + 子 agent 集群(10 个)

  • Director(调度官) — 选 idea、调 Research Team、决定 keep/revised/refute
  • Research Team(10 个子 agent 串行): IdeaCloser → LitSurveyer → MethodDesigner → ExpDesigner → PaperWriter → CitationGrounder → SectionValidator → CodeSketcher → Reviewer → Refuter
  • 每步任务极窄:IdeaCloser 只产 1 段 claim + 1 个假设;SectionValidator 只给 8 个 section 打 1-5 分

2️⃣ 5 源论文库(k 越高越丰富)

源 覆盖 默认
arXiv 物理 / CS / ML / quant bio 预印本 30+ query 并发
OpenAlex 跨学科 2.5 亿论文(最广覆盖) 4 query
bioRxiv 生物学预印本(microbiome 主战场) 4 query
PubMed/PMC 临床 + 基础生物 36M+(含 Cell/Nature/Lancet) 4 query
Semantic Scholar 1.95 亿论文 + 引文图(含 ICML/NeurIPS/CVPR) 4 query

跨源 **dedup** 走 3 级优先级:DOI > arXiv-id > 标题 hash

3️⃣ Claude-Code 风格终端 UI(Python `rich` 实现)

  • ASCII art banner
  • ● Tool name + ⎿ result 缩进(Claude Code 标志性样式)
  • ⠋ dots spinner
  • verdict 颜色表格(kept 绿 / revised 黄 / refuted 红)
  • 实时 status bar
  • REPL:❯ /help /status /ideas /vault /run /mass /demo /exit

🛞 两种运行模式

A. 单次 demo(5 个 idea,约 12-15 分钟)

python run_demo.py

B. Autopilot 连续自循环(生产模式)

# 默认 5 个 seed + 持续从 KB 挖新 idea
python run_autopilot.py

# 限制版本
python run_autopilot.py --max-cycles 2 --max-ideas