菜单
Deep-dive a single paper — methodology, assumptions, implications. Triggers on "read this paper", "读论文", "explain this PDF", "arxiv.org/abs/", paper URLs, "what does this paper say", "paper deep-dive".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Unified document archival — covers both agent-facing knowhow (env setup, debug, infra, runbooks) and human-facing docs (design, guides, README, changelog). Triggers on: "update docs", "更新文档", "写 README", "记下来", "归档", "save this", "update knowhow", or after a <knowhow-hint> prompt. Also triggers when conversation context implies documentation should be created or updated.
Alias for /update-docs (knowhow branch). Use when user says '记下来', '归档', 'save this', or after a <knowhow-hint> prompt. Routes to the unified /update-docs workflow with knowhow branch auto-selected.
Use when an experiment has completed and results need analysis, interpretation, or presentation. Triggers on "analyze results", "分析实验", "interpret experiment", "what did exp find", "run analysis".
Use when creating git commits, updating CHANGELOG.md, committing across nested repos, or generating weekly progress summaries. Triggers on "commit", "提交", "changelog", "weekly summary", "周报", "write commit message".
Use when setting up a new project (general or research) or when .pipeline-state.json is missing. Triggers on "init project", "初始化项目", "set up labmate", "first time setup", "initialize".
Check experiment status and diagnose failures via exp-manager. Triggers on "check status", "看看实验", "how's the run", "any failures", "is it done", "monitor experiment", "experiment status".
| name | read-paper |
| description | Deep-dive a single paper — methodology, assumptions, implications. Triggers on "read this paper", "读论文", "explain this PDF", "arxiv.org/abs/", paper URLs, "what does this paper say", "paper deep-dive". |
| disable-model-invocation | true |
Deep-dive a single paper with interactive Q&A. Outputs a structured analysis, then lets the user ask follow-up questions.
When this skill is invoked with <input>:
Determine input type and fetch accordingly:
| Input type | Detection | Action |
|---|---|---|
| Local PDF | <input> ends with .pdf and is a file path | Use Read tool. If PDF exceeds 20 pages, read pages 1-20 first, then ask user which sections to focus on. |
| arXiv PDF URL | <input> matches arxiv.org/pdf/ | Convert to abstract URL (replace /pdf/ with /abs/, remove .pdf suffix), then fetch with Jina Reader: curl -s "https://r.jina.ai/{abstract_url}" -H "Accept: text/markdown" |
| arXiv URL | <input> matches arxiv.org (non-PDF) | Fetch with Jina Reader: curl -s "https://r.jina.ai/{url}" -H "Accept: text/markdown" |
| Other URL | <input> starts with http:// or https:// | Try WebFetch first. If it fails or returns insufficient content, fall back to Jina Reader: curl -s "https://r.jina.ai/{url}" -H "Accept: text/markdown" |
| Pasted text | None of the above | Use <input> directly as paper content |
If fetching fails entirely, tell the user and suggest an alternative input method.
Read the following files if they exist (skip silently if not found):
docs/papers/landscape.md — user's literature mapexp/summary.md — user's experiment historyInvoke @domain-expert with this prompt:
Mode 4: Paper Deep-Dive
Paper content: {fetched paper content}
Research context:
- Literature landscape: {landscape.md content or "not available"}
- Experiment history: {exp/summary.md content or "not available"}
Provide your structured deep-dive analysis (Methodology Skeleton, Assumptions & Limitations, Bridge Analysis), then enter interactive Q&A mode.
The user will then interact directly with domain-expert for follow-up questions. When the user says "save" / "archive" / "store" / "存档" / "保存" (with optional "as {name}"), domain-expert handles archival via its Mode 4 archive sub-flow.