paper-intelligence

name	paper-intelligence
description	AI research paper intelligence — router skill that detects user intent and delegates to the appropriate workflow. Triggers on paper, arxiv, digest, research, 论文, arXiv IDs, ML method names, or academic research questions.

Paper Intelligence Router

Detect user intent and delegate to the right workflow skill or direct MCP tool call.

Persona Detection

Call paper_workspace_context() first. Read the mode field:

• "workspace" → user has reading history; show progress, suggest groups, auto-mark
• "lightweight" → just present results, skip workspace features

Intent Routing

User says	Route to
"start my day" / "今天看什么" / "morning"	paper_morning_brief tool directly
Gives a paper title / arXiv ID	paper_show or paper_find_and_download
"这篇论文讲了什么" / "analyze this"	deep-dive skill
"这个方向有什么工作" / "survey" / "综述"	literature-survey skill
"引用链" / "谁引了这篇" / "citation"	citation-explore skill
"筛一下" / "triage" / "哪些值得看"	paper-triage skill
"趋势" / "trend" / "这个方向火不火"	research-insight skill
Research question (e.g. "RL做placement还有没有新意？")	paper_research tool directly
"这篇可信吗" / "能不能复现" / "credibility"	paper_credibility tool directly
"跟踪这个方向" / "watch" / "关注这个作者"	paper_watch tool directly
"我觉得这篇不行" / "这类少推" / feedback	paper_feedback tool directly
"这篇能给我什么启发" / "research ideas"	research-planning skill (ideation mode)
"实验计划" / "能怎么复现改进" / "experiment plan"	research-planning skill (experiment mode)
"给我组个 reading pack" / "怎么读"	research-planning skill (reading pack mode)
"保存搜索结果" / "存一下这个列表"	paper_save_report tool directly
"推荐和我课题相关的" / "recommend"	paper_recommend tool directly
"抽取结构化信息" / "extract" / "比较表格"	paper_extract or paper_compare_table
"看看保存的报告" / "有哪些综述" / "list reports"	paper_list_reports tool directly
"我的偏好" / "我喜欢什么方向" / "preferences"	paper_preferences tool directly
"最近有新论文吗" / "watchlist 有更新吗"	paper_watch_check tool directly
"我关注了什么" / "watchlist"	paper_watch_list tool directly
"这篇的表格" / "提取表格" / "tables"	paper_tables tool directly
"哪些论文用了 GNN" / "query profiles"	paper_query tool directly
"我的阅读进度" / "reading stats"	paper_reading_stats tool directly
"看看我对这篇的笔记" / "show notes"	paper_note_show tool directly
"工作区概览" / "workspace status"	paper_workspace_status tool directly
Direct search keywords	paper_search or paper_quick_scan

Interaction Rules

• Context carry-over: When the user's request relates to papers already in this conversation:
  • Explicit reference ("根据已有的", "用刚才的", "基于这些"): Use those papers directly. No confirmation, no re-search.
  • Ambiguous (same-topic request, but no explicit reference): ASK "刚才找到了 N 篇相关论文，直接用这些？还是再补充搜索？"
  • New topic: Treat as new search, ignore context.
  • No context: Search directly.
• Intent-driven: When the user's intent is clear, skip intermediate steps and go straight to results. Don't ask clarifying questions unless intent is genuinely ambiguous.
• When routing to a tool directly (morning_brief, auto_triage, citation_trace), call it and present results. Then follow the Next-Step Prediction rules below.
• When routing to a skill, follow that skill's fork-only checkpoints.
• Never list more than 3 options. Prefer smart default + "或者？"
• Workspace operations (note_add, reading_status, group_add) are automatic. Report saving via paper_save_report is automatic in each workflow. Additional export (BibTeX) is opt-in via each workflow's FORK.

⚡ Next-Step Prediction Engine (MANDATORY for all tool/skill completions)

Every time a tool call or skill completes, you MUST proactively suggest 2-3 context-aware next actions. This is the core UX differentiator — the system anticipates what the researcher needs next.

Prediction Rules

After ANY tool/skill completion, analyze the result data and generate personalized suggestions:

Just completed	Result signals	Suggest
paper_search / paper_quick_scan	Found N papers	"找到 {N} 篇。要筛选一下（triage）？还是直接看评分最高的 [{top_title}]？"
paper_search	0 results	"本地没找到。要在线搜索？还是换个关键词？"
paper_morning_brief	Has important papers	"今天有 {N} 篇高相关。先看 [{top_title}]（{score}分）？还是先筛选一轮？"
paper_morning_brief	Empty digest	"今天没有新论文。要扫描一下 [{user_topic}] 的最新进展？"
paper_show	Paper loaded	"要深入分析这篇？看引用链？还是找类似的论文？"
paper_auto_triage	Has important bucket	"筛出 {N} 篇重要论文。要批量下载 PDF？还是先深入看 [{top_important_title}]？"
paper_auto_triage	All skip	"这批论文跟你方向关联不大。要调整 profile？还是换个方向搜索？"
paper_citation_trace	Found key nodes	"发现 {N} 个关键节点。要把这些加到阅读分组？还是继续追踪 [{key_node_title}]？"
paper_credibility	Low confidence	"这篇可信度偏低（{reason}）。要找同方向更可靠的替代论文？"
paper_credibility	High confidence	"可信度高。要深入分析方法细节？还是看实验能不能复现？"
paper_extract	Profile extracted	"结构化信息已提取。要跟其他论文对比？还是基于这篇做实验计划？"
paper_compare / paper_compare_table	Comparison done	"对比完成。要生成综述？还是导出 BibTeX 写 related work？"
paper_feedback	Feedback recorded	"已记录。后续推荐会参考你的偏好。要继续看下一篇？"
paper_recommend	Has recommendations	"推荐了 {N} 篇。要筛选一下？还是直接看 [{top_title}]？"
paper_watch	Watch added	"已关注。要看这个方向的趋势分析？"
paper_find_and_download	PDF downloaded	"PDF 已下载。要解析全文做深度分析？"
paper_research	Answer generated	"要把涉及的论文做个对比？还是深入某篇？"

Suggestion Format

Always use this format at the end of every response:

---
💡 **下一步建议**：
1. [最可能的下一步 — 基于当前结果数据]
2. [第二可能的下一步]
（直接说编号或描述你想做的事）

Key Principles

1. Data-driven: Suggestions must reference actual data from the result (paper titles, scores, counts). Never give generic suggestions.
2. Max 3 options: Never overwhelm. 2 is ideal, 3 is max.
3. Default action first: The most likely next action should be option 1.
4. Carry context: When user picks an option, carry all relevant paper IDs, scores, and context forward. Never ask the user to re-specify.
5. Exit is implicit: Don't add "还是先这样？" as an option — if the user wants to stop, they'll just stop or say something new.