DeepScan Monitor
Use this skill when the user wants Claude to manage a longer-running PapersFlow research workflow instead of a single search call.
Workflow
- 1. Use
run_deepscan to start the job. - Immediately tell the user that the run is asynchronous.
- Poll with
get_deepscan_live_snapshot for the best live view of:
- progress
- status message
- checkpoint state
- top papers
- partial summary
- key findings
- 4. Fall back to
get_deepscan_status if the user only wants lightweight progress checks. - Once
finalReportAvailable is true or the run is completed, call get_deepscan_report. - Use
summarize_evidence when the user wants a cross-report summary from stored DeepScan history. - Use
run_python_plot only after you have stable report data worth plotting.
Important Behavior
- - Do not imply the MCP server will push completion notifications into Claude automatically.
- Poll deliberately and explain that the run is being checked.
- Prefer
get_deepscan_live_snapshot over get_deepscan_status when the user wants richer live information. - If a report is not ready yet, say that clearly and keep the next action obvious.
Progress Update Style
When a run is still active, summarize:
- - current status
- progress percentage
- current stage or status message
- any checkpoint question
- notable live papers
- key findings if available
Keep updates brief unless the user asks for more detail.
Plotting Guidance
Use run_python_plot only for meaningful visualizations after you have stable report outputs, for example:
- - papers by year
- citation distribution
- venue distribution
- grouped comparison across a small number of finished runs
Do not generate plots for sparse or obviously low-quality data without saying so.
Examples
- - User asks: "Run a DeepScan on evaluation benchmarks for agentic retrieval systems and keep me posted."
- User asks: "Check how my DeepScan is progressing and tell me the key findings so far."
- User asks: "The run is finished, summarize the final report and plot papers by year."
- User asks: "Summarize the evidence from my recent DeepScan reports on protein structure prediction."
DeepScan Monitor
当用户希望Claude管理一个运行时间更长的PapersFlow研究工作流程，而非单次搜索调用时，请使用此技能。
工作流程
- 1. 使用rundeepscan启动任务。
- 立即告知用户该运行为异步执行。
- 使用getdeepscanlivesnapshot轮询获取最佳实时视图，包括：
- 进度
- 状态信息
- 检查点状态
- 顶级论文
- 部分摘要
- 关键发现
- 4. 如果用户仅需轻量级进度检查，则回退使用getdeepscanstatus。
- 当finalReportAvailable为true或运行完成后，调用getdeepscanreport。
- 当用户需要从存储的DeepScan历史记录中获取跨报告摘要时，使用summarizeevidence。
- 仅在获得值得绘图的稳定报告数据后，使用runpython_plot。
重要行为
- - 不要暗示MCP服务器会自动将完成通知推送到Claude。
- 有策略地进行轮询，并说明正在检查运行状态。
- 当用户希望获取更丰富的实时信息时，优先使用getdeepscanlivesnapshot而非getdeepscan_status。
- 如果报告尚未就绪，请明确说明，并保持下一步操作清晰可见。
进度更新风格
当运行仍在进行中时，总结以下内容：
- - 当前状态
- 进度百分比
- 当前阶段或状态信息
- 任何检查点问题
- 值得关注的实时论文
- 可用的关键发现
除非用户要求更多细节，否则保持更新简洁。
绘图指导
仅在获得稳定的报告输出后，使用runpythonplot进行有意义的可视化，例如：
- - 按年份统计论文
- 引用分布
- 发表场所分布
- 少量已完成运行的分组比较
不要为稀疏或明显低质量的数据生成图表，除非明确说明。
示例
- - 用户询问：对智能检索系统的评估基准运行一次DeepScan，并随时向我更新进度。
- 用户询问：检查我的DeepScan运行进度，并告诉我目前的关键发现。
- 用户询问：运行已完成，总结最终报告并按年份绘制论文分布图。
- 用户询问：总结我最近关于蛋白质结构预测的DeepScan报告中的证据。
