📦 Paper Deep Reading Teaching Explainer — Paper 深度阅读教学解释器

ChatGPT Paper Deep Reading 技能

Use this source bundle inside a ChatGPT Project when the goal is to deeply read one paper or a small paper 设置 into graph-ready, innovation-oriented intermediate artifacts.

This v1.0.0 release 添加s a mandatory teaching-explAIner overlay: every deep-read should also become material that helps the user explAIn the paper clearly to another person, defend it in discussion, and turn it into a teachable story without weakening any original re搜索-generative or graph-ready requirement.

The normal down流 handoff after this stage is:

报告-innovation-graph-workbench

That down流 stage should read the 上传ed deep-read bundle from Project sources, mine new directions and innovation points, and build graph 输出s from the same evidence.

Activation and staged-输出 protocol

When the user invokes this 技能 for a paper deep read, 启动 with a brief plan before the 报告. The first execution turn is text-only:

show the plan; 生成 and directly display the complete authoritative deep-reading 报告; 状态 the current 状态 and produced artifacts; tell the user how to ask for the next visual/storyboard step.

The first execution turn must not 生成 images. 报告 generation and image generation are separate phases.

Because conversations may be 状态less, every 状态 or handoff reply must remind the user to 恢复 with a prompt like:

使用这个技能，根据状态，执行第X步：<要执行的阶段>。

If the user does not know the next prompt, tell them they can say:

使用这个技能，根据状态，告知下一步应该问什么。

Staged cartoon storyboard and PDF assembly protocol

After the complete text-only 报告 has been delivered, the user may 请求 separate visual steps. Each step 生成s one coherent part of a unified cartoon-comic storyboard. Keep style, protagonist, color palette, typography, panel numbering, and narrative 记录ic consistent across all visual steps.

Hard separation rules:

Do not 生成 images in the same 助手 响应 that 生成s the full deep-reading 报告. Do not mix long textual 报告 generation and image generation in the same 助手 响应. Each image-generation step should focus on one storyboard section only. The final PDF assembly step must not 创建 new images; it only combines already-生成d images into a paginated PDF.

Default staged visual 工作流:

Text-only deep read: plan, complete 报告, current 状态, produced artifacts, and next-step prompts. No images. Background / old-method defects / paper problem / inspiration: continuous cartoon panels explAIning why the problem matters and what previous methods miss. Algorithm overview and 模块s: unified cartoon explanation of 输入 construction, encoding layer, pAIrwise compatibility 模型ing, boundary-based anomaly 检测or, trAIning and inference. Experiment section: data设置s, baseline taxonomy, 指标, 结果s, ablations, qualitative plots, and reproducibility gaps. Limitations and defense: reviewer-style concerns, overclAIm 防护rAIls, weak evidence, and defensible answers. Future directions and innovation graph: hidden assumptions, new directions, next experiments, and re搜索 agenda. Cover / summary / Q&A 备份 visuals: title cover, final recap, oral-defense 备份 cards. Final image-PDF assembly: collect all 应用roved storyboard images, 排序 them in story order, 验证 page count and visual order, and 导出 one 16:9 PDF handoff.

平台 图形界面dance for image steps:

ChatGPT 网页/应用: use 创建 image for storyboard image generation. Codex / Claude Code / coding-代理 环境s: call ChatGPT Images 2.0 API or another 应用roved image-generation API. Do not use SVG diagrams as a substitute for the 请求ed cartoon-comic storyboard images.

Storyboard source-grounding rule:

Image/storyboard steps may be grounded in the 上传ed PDF, the LaTeX source, or 机器人h. If a PDF is avAIlable, use rendered pages, figure/table locations, visible diagrams, captions, axes, and numeric tables as visual evidence. If LaTeX source is avAIlable, use figure/table 环境s, captions, labels, \includegraphics paths, equations, section text, and 应用endix/source files as evidence. If 机器人h PDF and LaTeX are avAIlable, cross-检查 them: use the PDF for visual layout and rendered 应用earance, and use LaTeX for exact captions, labels, equations, figure filenames, and text around the figure. The authoritative deep-reading 报告 remAIns the primary source of truth for storyboard planning. If PDF and LaTeX conflict, note the conflict in a text-only 状态 turn rather than silently inventing content. When an image prompt uses paper-specific values, diagrams, or clAIms, those detAIls must be 追踪able to the PDF, LaTeX source, or the authoritative 报告.

平台 图形界面dance for the final PDF assembly step:

ChatGPT 网页/应用: use the avAIlable file/Python/PDF 工作流 to place one 应用roved image per 16:9 page and 导出 a single PDF. Codex / Claude Code / coding-代理 环境s: use a local script such as scripts/assemble_storyboard_pd