Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →**技能描述**: 基于 Obsidian 的 L3 级论文阅读与记忆抽取系统
.claude/skills/eesjgong-scholarskill/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | 109% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 124% | 0% |
| case-01 | ✗→✓ | ▲ Improved | 53% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 146% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 147% | 0% |
技能描述: 基于 Obsidian 的 L3 级论文阅读与记忆抽取系统
版本: v1.0.0 最后更新: 2026-03-18
欢迎使用 ScholarSkill！ 🦉
在开始阅读论文之前，需要完成以下配置。本技能支持三种配置模式：
bash# 先选择中文包，再明确指定目标 Obsidian vault python ~/.openclaw/workspace-scholar/skills/scholar-skill/zh-CN/scripts/configure.py auto --vault-path "/你的/Obsidian/Vault"
适合: 第一次使用，Obsidian 已安装 自动完成:
注意: 推荐始终传入 --vault-path，这样不会误选到别的 vault，也能确保只同步中文模板。
bash# 运行中文包配置脚本，并明确指定目标 vault python ~/.openclaw/workspace-scholar/skills/scholar-skill/zh-CN/scripts/configure.py semi --vault-path "/你的/Obsidian/Vault"
适合: 希望手动确认每一步的用户 流程:
bash# 运行中文包配置脚本，并明确指定目标 vault python ~/.openclaw/workspace-scholar/skills/scholar-skill/zh-CN/scripts/configure.py manual --vault-path "/你的/Obsidian/Vault"
适合: 清楚自己需要什么的高级用户 流程:
本技能需要以下依赖技能（配置脚本会自动检查）：
| 技能 | 用途 | 必需 | |------|------|------| | obsidian-direct | Obsidian 文件操作 | ⭐⭐⭐ | | arxiv-watcher | ArXiv 论文搜索 | ⭐⭐⭐ |
| 技能 | 用途 | 必需 | |------|------|------| | academic-research-hub | 多源学术搜索 | ⭐⭐ | | tavily | 网页内容提取 | ⭐⭐ | | pdf | PDF 文本提取 | ⭐⭐ | | obsidian-cli | Obsidian CLI 工具 | ⭐⭐ |
| 技能 | 用途 | 必需 | |------|------|------| | durable-task-runner | 长任务编排、进度追踪、崩溃恢复 | ⭐⭐⭐ (L3/批量) |
为什么需要 durable-task-runner？
安装命令:
bash# 核心依赖（通过 ClawHub 安装） clawhub install obsidian-direct clawhub install arxiv-watcher # 功能增强（推荐） clawhub install academic-research-hub clawhub install tavily clawhub install pdf clawhub install obsidian-cli # 长任务编排（L3/批量处理必需） clawhub install durable-task-runner
手动安装（如 ClawHub 不可用）:
bash# 从 GitHub 克隆技能 cd ~/.openclaw/workspace-scholar/skills git clone https://github.com/OpenClaw/obsidian-direct.git git clone https://github.com/OpenClaw/arxiv-watcher.git # ... 其他技能
配置完成后，配置文件位于：
~/.openclaw/workspace-scholar/config/scholar.yml（如该目录不存在，则写入 ~/.openclaw/workspace/config/scholar.yml）
关键配置项:
yamlobsidian: vault_path: /Users/shengjie/ObsidianVault # ← 自动检测或手动输入 reading: default_level: L2 # L1/L2/L3 notification: feishu_enabled: false # 是否启用飞书推送
Q: 找不到 Obsidian 仓库怎么办？ A: 使用半自动或手动模式，手动输入路径。如未安装 Obsidian，先下载安装：https://obsidian.md
Q: 依赖技能安装失败？ A: 使用 ClawHub 安装：clawhub install <技能名>。如 ClawHub 不可用，从 GitHub 手动克隆技能到 ~/.openclaw/workspace-scholar/skills/ 目录。
Q: 可以修改配置吗？ A: 可以！编辑 ~/.openclaw/workspace-scholar/config/scholar.yml（或 ~/.openclaw/workspace/config/scholar.yml）
Q: 配置错了想重新配置？ A: 删除配置文件后，用中文包重新运行：rm ~/.openclaw/workspace-scholar/config/scholar.yml && python ~/.openclaw/workspace-scholar/skills/scholar-skill/zh-CN/scripts/configure.py auto --vault-path "/你的/Obsidian/Vault"
Semantic Memory: 事实、概念、方法、结论 Episodic Memory: 疑问、误解、修正过程 Procedural Memory: 可复用的阅读规则和研究方法
0-Inbox/1. 接收论文 (PDF/ArXiv URL/本地文件)
↓
2. 优先级评估 (P0/P1/P2)
↓
3. 选择阅读级别 (L1/L2/L3)
↓
4. 执行阅读 + 笔记生成
↓
5. 记忆抽取 (Semantic/Episodic/Procedural)
↓
6. 更新 2-Knowledge/（Concept / Insight / Method / Question / Person）
↓
7. 知识关联 (双向链接 + MOC 更新)
↓
8. 反思 / 确认 / 输出报告yaml# ~/.openclaw/workspace-scholar/config/scholar.yml 或 ~/.openclaw/workspace/config/scholar.yml # 注意：配置脚本优先写入 ~/.openclaw/workspace-scholar/config/（如该目录存在） obsidian: inbox_folder: 0-Inbox vault_path: /Users/shengjie/ObsidianVault paper_notes_folder: 1-Papers knowledge_folder: 2-Knowledge concepts_folder: 2-Knowledge/Concepts insights_folder: 2-Knowledge/Insights methods_folder: 2-Knowledge/Methods questions_folder: 2-Knowledge/Questions people_folder: 2-Knowledge/People moc_folder: 3-MOCs outputs_folder: 4-Outputs reflections_folder: 4-Outputs/Reflections confirmation_records_folder: 4-Outputs/Confirmation-Records templates_folder: 9-Templates/zh-CN memory_folder: memory reading: default_level: L2 enable_memory_extraction: true enable_knowledge_consolidation: true notification: feishu_enabled: false feishu_user_id: ou_xxxxxxxxxxxxx
配置脚本会把中文模板同步到：
text{vault}/9-Templates/zh-CN/
位置：{vault}/1-Papers/By-Topic/{Topic}/{Year}-{Author}-{Title}.md
位置：{vault}/memory/semantic/{Topic}.md
位置：{vault}/memory/procedural/{Topic}.md
位置：{vault}/memory/episodic/{Date}-{Paper}.md
位置：{vault}/3-MOCs/MOC-{Topic}.md
位置：{vault}/2-Knowledge/Concepts/
位置：{vault}/2-Knowledge/Insights/
位置：{vault}/2-Knowledge/Questions/
位置：{vault}/2-Knowledge/Methods/
位置：{vault}/2-Knowledge/People/
位置：{vault}/4-Outputs/Reflections/L1|L2|L3/
位置：{vault}/0-Inbox/
用户：请用 L2 级别阅读这篇论文
附件：paper.pdf
智能体:
1. 评估优先级 → P1
2. 执行 L2 阅读（45 分钟标准）
3. 生成笔记（3-5KB）
4. 抽取 5-8 条记忆
5. 需要时创建 Concept / Insight / Question / Method 卡片
6. 更新知识关联与 MOC
7. 生成 L1 反思
8. 输出报告用户：请以 L3 级别精读 ArXiv:2407.19354
智能体:
1. 获取论文（ArXiv API）
2. 评估优先级 → P0（与核心方向直接相关）
3. 执行 L3 阅读（2.5 小时标准）
4. 生成深度笔记（10-15KB）
5. 知识升级 + 旧知识修订
6. 提炼 2-3 条程序规则
7. 更新 2-Knowledge/ 与 3-MOCs/
8. 若出现冲突/新方向，创建确认请求到 0-Inbox/
9. 生成深度反思并推送详细报告用户：这 10 篇论文请先 L1 筛选，然后对其中的 P0/P1 论文执行 L2 阅读
智能体:
1. 批量 L1 筛选（5 分钟/篇）
2. 评级分类：P0(2 篇) + P1(5 篇) + P2(3 篇)
3. 对 P0 论文执行 L3 阅读
4. 对 P1 论文执行 L2 阅读
5. P2 论文仅存档
6. 周末执行 L2 反思并整理概念 / MOC
7. 输出汇总报告arxiv-watcher: ArXiv 论文搜索academic-research-hub: 多源学术搜索obsidian-direct: Obsidian 文件操作tavily: 网页内容提取pdf: PDF 文本提取obsidian-cli: Obsidian CLI 工具contradict，等待周巩固时裁决技能作者: Scholar Agent (学究) 最后更新: 2026-03-18 许可证: MIT
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-10 | fail→pass | 18,648 | 8,608 | -54% | 1 | 1 | 0% | 2,439 | 5,090 | +109% | 0 | 0 | — |
case-09 | fail→pass | 17,868 | 7,569 | -58% | 1 | 1 | 0% | 2,399 | 5,378 | +124% | 0 | 0 | — |
case-01 | fail→pass | 23,651 | 23,584 | -0% | 1 | 1 | 0% | 4,336 | 6,637 | +53% | 0 | 0 | — |
case-02 | fail→fail | 49,968 | 12,764 | -74% | 1 | 1 | 0% | 8,154 | 4,649 | -43% | 0 | 0 | — |
case-03 | fail→fail | 11,585 | 7,798 | -33% | 1 | 1 | 0% | 1,963 | 5,228 | +166% | 0 | 0 | — |
case-04 | fail→pass | 21,352 | 21,249 | -0% | 1 | 1 | 0% | 2,963 | 7,279 | +146% | 0 | 0 | — |
case-05 | fail→pass | 14,989 | 10,733 | -28% | 1 | 1 | 0% | 2,250 | 5,565 | +147% | 0 | 0 | — |
case-06 | fail→pass | 17,591 | 8,928 | -49% | 1 | 1 | 0% | 2,658 | 5,176 | +95% | 0 | 0 | — |
case-07 | fail→fail | 13,706 | 8,734 | -36% | 1 | 1 | 0% | 2,042 | 5,187 | +154% | 0 | 0 | — |
case-08 | fail→pass | 21,691 | 16,640 | -23% | 1 | 1 | 0% | 3,263 | 6,055 | +86% | 0 | 0 | — |
case-11 | fail→pass | 13,778 | 4,876 | -65% | 1 | 1 | 0% | 1,879 | 4,755 | +153% | 0 | 0 | — |
case-12 | fail→pass | 14,994 | 7,325 | -51% | 1 | 1 | 0% | 2,066 | 5,210 | +152% | 0 | 0 | — |
case-13 | fail→pass | 14,549 | 7,248 | -50% | 1 | 1 | 0% | 2,317 | 4,988 | +115% | 0 | 0 | — |
case-14 | fail→pass | 15,584 | 6,103 | -61% | 1 | 1 | 0% | 1,857 | 4,995 | +169% | 0 | 0 | — |
case-15 | pass→pass | 13,733 | 4,696 | -66% | 1 | 1 | 0% | 2,417 | 4,773 | +97% | 0 | 0 | — |
case-16 | pass→pass | 17,311 | 15,107 | -13% | 1 | 1 | 0% | 2,573 | 6,211 | +141% | 0 | 0 | — |
case-17 | fail→pass | 21,052 | 16,381 | -22% | 1 | 1 | 0% | 3,168 | 6,535 | +106% | 0 | 0 | — |
case-18 | fail→fail | 25,229 | 19,912 | -21% | 1 | 1 | 0% | 3,152 | 6,694 | +112% | 0 | 0 | — |
case-19 | fail→pass | 13,821 | 2,311 | -83% | 1 | 1 | 0% | 1,815 | 4,251 | +134% | 0 | 0 | — |
case-20 | fail→pass | 5,823 | 7,922 | +36% | 1 | 1 | 0% | 845 | 5,035 | +496% | 0 | 0 | — |
case-21 | fail→fail | 7,529 | 8,095 | +8% | 1 | 1 | 0% | 1,158 | 4,995 | +331% | 0 | 0 | — |
case-22 | pass→pass | 24,860 | 35,091 | +41% | 1 | 1 | 0% | 3,707 | 8,500 | +129% | 0 | 0 | — |
case-23 | fail→pass | 11,875 | 3,474 | -71% | 1 | 1 | 0% | 1,524 | 4,365 | +186% | 0 | 0 | — |
case-24 | fail→pass | 19,197 | 8,777 | -54% | 1 | 1 | 0% | 2,786 | 5,447 | +96% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 24 cases were attempted, and 23 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +67 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.