Huo15 Research Pipeline

从想法到论文的全自主研究管道，6阶段，含HITL，输出到Obsidian

Security Scan

The skill is internally consistent with its stated goal (an LLM-driven multi‑phase research pipeline that saves outputs to the user's Obsidian/agent KB); it uses only the OpenClaw CLI and standard tools and does not request unrelated credentials or global privileges.

Purpose & Capability

Name/description (research pipeline → paper output to Obsidian) match the actual code and runtime instructions. The scripts call the OpenClaw CLI for LLM generation and web search, use curl/jq/obsidian optionally — these are reasonable for automated literature search, LLM composition, and optional Obsidian sync. File writes are scoped to $HOME/.openclaw/agents/.../research-..., which is consistent with storing generated research artifacts.

Instruction Scope

Runtime instructions and scripts perform the six phases described, prompt the LLM with aggregated local phase outputs, and save stage artifacts. The script invokes 'openclaw llm generate' and 'openclaw web search' (network calls) and writes outputs to disk. This is expected, but users should note that prompts include their topic and previous phase contents — that data will be sent to the configured LLM/web search service. The script does not read arbitrary system files or request additional environment variables beyond relying on the user's OpenClaw configuration.

Install Mechanism

There is no install spec (instruction-only install), so nothing is downloaded or installed by the skill itself. The provided code is shell scripts that run locally; no external download/extract URLs or non-standard installers are present.

Credentials

The skill declares no required env vars or credentials, which is appropriate. It does implicitly depend on the user's OpenClaw CLI configuration (which may contain API keys/tokens for LLM providers/search backends). That dependency is proportional to its function, but users should be aware their research text will be sent to whatever model/service OpenClaw is configured to use.

Persistence & Privilege

always is false and the skill does not attempt to modify other skills or global agent settings. It creates and writes files under the agent KB/research directory and optionally calls an 'obsidian' CLI if available. Those behaviors are within the expected scope for storing outputs and syncing to Obsidian.

Assessment

This skill appears to do what it says: run six research phases using your local OpenClaw CLI and save outputs to an agent KB folder (and optionally Obsidian). Before running it: (1) Inspect the scripts (you already have them) and confirm you are comfortable with files being written to $HOME/.openclaw/agents/...; (2) Remember that prompts and intermediate outputs are sent to whatever LLM/web-search service your OpenClaw CLI is configured to use — do not submit sensitive or proprietary data unless you trust that service; (3) If you use Obsidian sync, check what the 'obsidian' CLI will do and whether it requires credentials; (4) Run the script first with a harmless test topic to observe network calls and outputs; (5) Consider running in an isolated account/container if you want to limit blast radius. Overall there are no unexplained credential requests or privileged operations in the skill.

从想法到论文的全自主研究管道，灵感来自 AutoResearchClaw

元信息

• version: 1.0.0
• trigger: 研究管道、research pipeline、自动研究、想法到论文、研究 [课题]
• author: huo15
• compatibility: OpenClaw >= 1.0

功能概述

将一个研究课题全自动推进至论文产出，涵盖范围定义、文献发现、知识综合、实验设计、结果分析与论文撰写。全程人类在环（HITL），每个 Phase 完成后等待用户确认。

核心流程

用户: "研究 [课题]"
         │
         ▼
  ┌─────────────────┐
  │  Phase A: 范围定义  │  → 用户确认
  └────────┬────────┘
           ▼
  ┌─────────────────┐
  │  Phase B: 文献发现  │  → 用户确认
  └────────┬────────┘
           ▼
  ┌─────────────────┐
  │  Phase C: 知识综合  │  → 用户确认
  └────────┬────────┘
           ▼
  ┌─────────────────┐
  │  Phase D: 实验设计  │  → 用户确认
  └────────┬────────┘
           ▼
  ┌─────────────────┐
  │  Phase E: 结果分析  │  → 用户确认
  └────────┬────────┘
           ▼
  ┌─────────────────┐
  │  Phase F: 论文撰写  │  → 完成
  └────────┬────────┘
           ▼
      输出到 Obsidian

Phase 详细说明

Phase A: 研究范围定义

定义研究问题、目标、范围边界和成功标准。

输出：

• 研究问题（1-3 个核心问题）
• 研究目标（ SMART 格式）
• 范围边界（ in-scope / out-of-scope）
• 成功标准

Phase B: 文献发现与筛选

搜索相关论文、博客、技术报告，筛选高质量来源。

输出：

• 关键词列表
• 筛选出的文献列表（含标题、摘要、链接）
• 文献质量评分

Phase C: 知识综合与假设生成

整合文献发现，生成假设或研究问题。

输出：

• 知识图谱摘要
• 核心发现列表
• 假设 / 待验证命题

Phase D: 实验设计与执行

设计验证假设的实验方案，包括数据、方法和评估指标。

输出：

• 实验设计方案
• 数据需求
• 评估指标

Phase E: 结果分析

分析实验结果，生成洞察。

输出：

• 结果摘要
• 统计显著性（如适用）
• 洞察列表

Phase F: 论文撰写

按学术论文结构输出完整内容。

输出：

• 标题 + 摘要
• 引言
• 相关工作
• 方法
• 结果
• 讨论
• 结论
• 参考文献

使用方式

用户: 研究 大语言模型在代码补全任务中的性能评估

或直接说：

用户: 自动研究 基于强化学习的机器人抓取策略

输出位置

所有研究产物保存至：

$HOME/.openclaw/agents/main/agent/kb/raw/research-{课题名}-{日期}/

结构：

research-{课题}-{日期}/
├── 00_scope.md        # Phase A 输出
├── 01_discovery.md    # Phase B 输出
├── 02_synthesis.md    # Phase C 输出
├── 03_experiment.md   # Phase D 输出
├── 04_analysis.md     # Phase E 输出
├── 05_paper.md        # Phase F 输出（完整论文）
└── log.txt            # 执行日志

人类在环（HITL）

每个 Phase 完成后，脚本会：

1. 显示该阶段产出摘要
2. 询问用户：继续下一步 / 调整参数 / 终止
3. 等待用户输入后执行下一阶段

调用示例

cd ~/.openclaw/workspace/skills/huo15-research-pipeline
./scripts/research.sh "大语言模型在代码补全任务中的性能评估"

依赖

• openclaw CLI（用于 LLM 调用）
• curl（用于 Web 搜索）
• jq（用于 JSON 解析）
• obsidian-cli（可选，用于直接写入 Obsidian）

注意事项

• 研究过程可能耗时较长，建议在空闲时运行
• 每个 Phase 的产出都会保存，中断后可从断点继续
• 论文撰写 Phase 会尽量保持学术规范，可直接提交或进一步编辑