Paper Digest
Input
- -
arxiv_id: bare ID like 2305.11206 or full https://arxiv.org/abs/2305.11206 (Normalise: strip the URL prefix, extract bare ID.)
Step 1 — Fetch the main paper
Try HTML first: web.fetch https://arxiv.org/html/
- - If HTTP 200 → use this as
paper_text. - If Falis: add v1 at the end, and try again: web.fetch https://arxiv.org/html/v1
- else if all fails skip and ABORT!
If the paper_text is retrieved then write the summary to ~/.openclaw/workspace/papers/<arxiv-id>.md.
Step 2 — Extract citations
Note: DO NOT DO THIS STEP INSIDE sub-agents
Within the main agent and from paper_text, identify at most 5 citations the paper most directly builds on. Prioritize:
- - Papers that are explicitly extended or improved upon
- Papers used as the primary baseline for comparison
- Papers that provide the core architecture this work adopts
- Papers referred to repeatedly (not just mentioned once) or that provide essential context
For each citation, extract either the arXiv ID or the title. Then resolve to an arXiv URL:
- - If an arXiv ID is in the reference → INLINECODE9
- Otherwise search
https://arxiv.org/search/?query=<title>&searchtype=all and take the first match.
Step 3 — Spawn sub-agents for citations
Note: Ensure that the sub-agent related task is precise and concise so the sub-agent does not have to re-read the previously read SKILLs and files.
For each resolved citation:
- - Check if the file for citation exists:
~/.openclaw/workspace/papers/<arxiv-id>.md, if it does then skip and consider the sub-agent concluded. - If the previous step fails then spawn a sub-agent with this EXACT instruction in VERBATIM:
- Fetch https://arxiv.org/html/
id> (or add v1 at the end, and try again: web.fetch https://arxiv.org/html/id>v1). If unavailable, SKIP. If retrieved then Write the summary to ~/.openclaw/workspace/papers/<arxiv-id>.md.
Step 4 — Write the executive summary
Check the citation summaries within ~/.openclaw/workspace/papers/ then utilize the main paper we are summarising with citation summaries and write a single markdown document in flowing prose (no bullet lists) to ~/.openclaw/workspace/digest/report_<arxiv-id>. Use this structure:
CODEBLOCK0
Rules
- - Every citation must be a markdown link: INLINECODE15
- No bullet lists in the output — prose only.
- If a section is absent from the paper (e.g., no ablations), skip it silently.
- Do not fabricate results, metrics, or author claims.
- Citation resolution retry: If a citation URL cannot be resolved after one retry, write the citation as plain text without a link:
[Author et al., YEAR].
技能名称: paper-digest
详细描述:
论文摘要
输入
- - arxiv_id: 裸ID如2305.11206或完整https://arxiv.org/abs/2305.11206(标准化:去除URL前缀,提取裸ID。)
步骤1 — 获取主论文
首先尝试HTML:web.fetch https://arxiv.org/html/
- - 如果HTTP 200 → 将此作为papertext。
- 如果失败:在末尾添加v1,然后重试:web.fetch https://arxiv.org/html/id>v1
- 否则如果全部失败,跳过并中止!
如果成功获取paper_text,则将摘要写入~/.openclaw/workspace/papers/.md。
步骤2 — 提取引用
注意:不要在子代理内部执行此步骤
在主代理内部,从paper_text中识别该论文最直接依赖的最多5篇引用。优先考虑:
- - 明确扩展或改进的论文
- 用作比较的主要基线的论文
- 提供本工作采用的核心架构的论文
- 反复提及(而非仅一次提及)或提供必要背景的论文
对于每篇引用,提取arXiv ID或标题。然后解析为arXiv URL:
- - 如果引用中包含arXiv ID → https://arxiv.org/html/
- 否则搜索https://arxiv.org/search/?query=&searchtype=all并取第一个匹配项。
步骤3 — 为引用生成子代理
注意:确保子代理相关任务精确简洁,以便子代理无需重新阅读之前已读的技能和文件。
对于每个解析出的引用:
- - 检查引用文件是否存在:~/.openclaw/workspace/papers/.md,如果存在则跳过并视子代理任务完成。
- 如果上一步失败,则生成一个子代理,并逐字使用以下精确指令:
- 获取https://arxiv.org/html/
id>(或在末尾添加v1,然后重试:web.fetch https://arxiv.org/html/id>v1)。如果不可用,跳过。如果获取成功,则将摘要写入~/.openclaw/workspace/papers/.md。
步骤4 — 撰写执行摘要
检查~/.openclaw/workspace/papers/中的引用摘要,然后利用我们正在总结的主论文及引用摘要,撰写一份流畅散文形式的单一markdown文档(无项目符号列表),保存至~/.openclaw/workspace/digest/report_。使用以下结构:
markdown
<标题>
<该问题解决什么及其重要性。背景及相关参考文献摘要>
<先前工作的不足以及本文如何弥补这一差距。内联引用为作者等,年份。
<核心方法,用通俗语言描述>
<主要结果。与先前工作的差异>
<一个局限性和一个未来方向>
<详细的消融实验、基准测试(如果本文或引用参考文献中存在)>
规则
- - 每篇引用必须是markdown链接:作者等,年份
- 输出中无项目符号列表 — 仅限散文。
- 如果论文中缺少某个部分(例如,无消融实验),则静默跳过。
- 不得捏造结果、指标或作者声明。
- 引用解析重试:如果引用URL在一次重试后仍无法解析,则将引用写为无链接的纯文本:[作者等,年份]。