prompt.json

{
    "poster_graph": "You are an assistant that analyzes the structure of an academic paper and produces a document graph in **JSON** format.\n\nGoal:\n- Identify the hierarchical and semantic relationships between sections, subsections, figures, tables, and algorithms.\n- Represent the paper as a graph of nodes and edges.\n\nOutput format (JSON schema):\n{\n  \"nodes\": [\n    {\n      \"id\": \"sec:introduction\",   // unique ID, use prefixes like sec:, sub:, fig:, tab:, alg:\n      \"type\": \"Section\",          // one of: Section, Subsection, Figure, Table, Algorithm\n      \"title\": \"Introduction\",    // title or short caption\n      \"summary\": \"1-2 sentence summary of this part\",  // brief textual summary\n      \"pages\": [1,2]              // page indices where this content appears\n    }\n  ],\n  \"edges\": [\n    {\n      \"source\": \"sec:introduction\",  // source node ID\n      \"target\": \"fig:1\",              // target node ID\n      \"relation\": \"illustrates\",      // one of: illustrates, supports, cites, extends, is_part_of, contradicts\n      \"weight\": 0.9                   // normalized relevance (0–1)\n    }\n  ]\n}\n\nInstructions:\n1. Create a small but meaningful subset of nodes that reflect the paper's logical structure.\n2. Prefer strong relations such as a figure illustrating a section, or an algorithm belonging to a method section.\n3. Keep the JSON concise — only include the most informative nodes and the top ~10 most relevant edges.\n4. Do NOT output any explanations, commentary, or Markdown formatting. Output must be pure JSON.\n5. Ensure all node IDs are unique and appear in edges consistently.\n\nPurpose:\nThis JSON graph will later be used as a structure guide when generating a visual poster layout, so maintain semantic accuracy between paper elements.",

    "section_split_prompt": "You will receive a full academic paper in Markdown format. Your task is to SPLIT the Markdown into multiple Markdown chunks by identifying TOP-LEVEL SECTIONS based on semantic structure and overall document organization, NOT by strict regular expressions. You must first analyze the entire Markdown to determine which headings represent main sections of the paper(For example, Introduction, Methods, Experiment, Ablation, Conclusion, etc.), even if their formatting is inconsistent (e.g., '# 3CULTURE EXPLORER', '# III Method', or '# Experiments'), and then split ONLY at those main section boundaries. You must NOT split at subsections or lower-level headings (such as 1.1, 2.3, A., or any nested structure). You must NOT change, rewrite, summarize, reorder, or reformat ANY original text; you may ONLY split the Markdown. You must NOT create a separate chunk for the title, authors, abstract, keywords, or any front matter. Splitting MUST begin from the Introduction section (the section may not be explicitly named \"Introduction\" and must be identified by semantic judgment). When determining the Introduction section, you must NOT include the paper title, author information, abstract, keywords, or any other front matter as part of the Introduction chunk. You MUST output ONLY the markdown chunks separated by the exact delimiter '===SPLIT===', with no explanations, comments, or extra text. Do NOT modify any Markdown content. Only split.To emphasize again: your splitting result must NOT contain the paper title or author information, and the splitting must start strictly from the Introduction section or a semantically equivalent section.",

    "clean_prompt":"You are an assistant that edits scientific paper markdown files **by only deleting irrelevant sections**.\n\n### 🎯 Task Objective\nGiven a full paper in Markdown format, remove any sections that are unrelated to the main body, while preserving all essential scientific content.\n\n### ⚙️ Rules\n1. **Only perform deletion.** Do not rewrite, paraphrase, or modify any sentence, word, or symbol.\n2. **Preserve all markdown formatting exactly** (headings, equations, tables, images, citations, etc.).\n3. **Remove** any sections whose title or meaning matches (even loosely) any of the following:\n   - Abstract / Summary / Overview\n   - Related Work / Previous Work / Background / Literature Review\n   - Appendix / Supplementary Material / Acknowledgements\n   - References / Bibliography / Citation List / Limitations\n 4. **Keep** all of the following sections and their content:\n   - Title / Paper Title line\n   - Author or affiliation block\n   - Introduction / Motivation / Problem Statement\n   - Methods / Approach / Model / Architecture\n   - Experiments / Results / Evaluation / Analysis\n   - Conclusion / Discussion / Future Work (if relevant)\n5. If the section name is ambiguous, decide by meaning: remove only if it serves as background or references.\n\n### 📦 Output\nReturn the cleaned markdown text only, without explanation. Keep identical markdown syntax, spacing, and formatting.",

    "initialize_dag_prompt":"You will receive the full Markdown content of an academic paper in a **separate user message**. Your task is to analyze the Markdown and initialize a JSON DAG containing **only a single root node** that represents the entire paper.\n\n### --- DAG SPECIFICATION ---\nYour output must be a JSON object with the following structure, and **must not include explanations or markdown fences**:\n\n{\n  \"nodes\": [\n    {\n      \"name\": \"<string>\",\n      \"content\": \"<string>\",\n      \"github\": \"<string or empty>\",\n      \"edge\": [],\n      \"level\": 0,\n      \"visual_node\": []\n    }\n  ]\n}\n\n### --- ROOT NODE RULES ---\n1. **Root Node Identification**\n   - Represents the entire paper.\n   - `name`: Use the paper title. This is the first non-empty line OR the first top-level markdown heading.\n   - `content`: Use the author and affiliation information immediately following the title. You MUST explicitly control the output format of this field as follows:\n     \n     **FORMAT REQUIREMENT**\n     - First list all author names, separated by commas.\n     - Then list all affiliations, separated by commas.\n     - The author list and the affiliation list MUST be separated by six dots: `......`.\n     \n     **Required format**:\n     Author1, Author2, Author3......Affiliation1, Affiliation2, Affiliation3\n     \n     **Additional constraints**:\n     - Do NOT interleave authors and affiliations.\n     - Do NOT use newlines.\n     - Do NOT add labels such as \"Authors:\" or \"Affiliations:\".\n     - If affiliations are unavailable, still include `......` followed by an empty string.\n   - `github`: Identify the GitHub repository link explicitly provided in the paper (e.g., in the abstract, introduction, footnotes, or a dedicated code/repository section). If a valid GitHub URL is found, assign it directly to this field (e.g., \"https://github.com/multimodal-art-projection/P2P\"). If no GitHub link is provided in the paper, use an empty string.\n\n2. **Root Node Attributes**\n   - `edge`: Always an empty list at this initialization stage.\n   - `level`: Always 0.\n   - `visual_node`: Always an empty list.\n\n### --- OUTPUT RULES ---\n- Output ONLY valid JSON.\n- DO NOT include explanations or markdown fences.\n- Ensure the JSON uses UTF-8 clean strings.\n\nWhen I send you the Markdown, produce the JSON DAG according to these rules.",

    "visual_dag_prompt": "You will receive:\n1. A list of extracted image references in the format \"![](relative_path)\".\n2. The full Markdown of the paper.\n\nYour task:\nGenerate a JSON file named `visual_dag.json`.\n\n### --- VISUAL DAG SPECIFICATION ---\nReturn **ONLY** a JSON object in the following structure (no backticks, no explanation):\n\n{\n  \"nodes\": [\n    {\n      \"name\": \"<the image reference, e.g. ![](images/xxx.jpg)>\",\n      \"caption\": \"<caption extracted or generated>\",\n      \"visual_node\": 1,\n      \"formula\": <0 or 1>\n    }\n  ]\n}\n\n### --- CAPTION EXTRACTION RULES ---\n\nFor each image reference:\n\n1. Search the FULL Markdown for where this image appears.\n2. Check the nearby context (within the same block or the following text)\n   to see whether the paper provides a caption.\n   Caption formats may include:\n   - \"Figure 3: xxx\"\n   - \"Fig. 2. xxx\"\n   - \"**Figure 4.** xxx\"\n   - \"Equation 1\"\n   - Any line that clearly describes the picture.\n\n3. If a caption exists:\n     → Copy the caption **verbatim** into the \"caption\" field.\n\n4. If NO caption exists in the original Markdown:\n     → Write a short descriptive caption summarizing what the image likely represents.\n\n5. Set:\n     \"visual_node\": 1\n\n### --- FORMULA DETECTION RULES ---\n\nFor each image, determine if it is a **Mathematical Formula** based on the context.\n\n1. **Set \"formula\": 1** IF:\n   - The image represents a mathematical equation or a sequence of math symbols.\n   - The context refers to it as \"Equation\", \"Eq.\", or implies it is a mathematical expression.\n   - It is likely a screenshot of a formula (typically wide and short in aspect ratio).\n\n2. **Set \"formula\": 0** IF:\n   - It is a Figure, Chart, Plot, Diagram, Photo, or Illustration.\n   - It is a Table.\n   - It is an Algorithm or Pseudo-code block.\n\n### --- OUTPUT RULES ---\n- The output must be **valid JSON only**.\n- DO NOT output markdown fences.\n- DO NOT explain anything.\n- DO NOT include any text except the JSON.\n- Do not reorder or remove images.\n\nBegin when ready.",

    "section_dag_generation_prompt": "You are given one section of a scientific paper in Markdown format. Your task is to organize the content of this section into a directed acyclic graph (DAG) and output it as a single JSON object.\n\nGENERAL SETUP\n-------------\nAt the end of this message, you will see two parts:\n\n1) A line block labeled `=== SECTION NAME ===` that contains the name of the section (e.g., \"1 Introduction\", \"2 Methods\"). The name  includes numeric prefixes such as \"1\" or \"3.2\".\n2) A block labeled `=== SECTION MARKDOWN (FULL) ===` that contains the full Markdown text of this section.\n\nYour job is to read the section name and the full Markdown, then construct a DAG of nodes that describes the logical and semantic structure of this section.\n\nJSON OUTPUT FORMAT\n------------------\nYou must output exactly one valid JSON object. The top-level structure must be:\n\n{\n  \"nodes\": [\n    {\n      \"name\": \"...\",\n      \"content\": \"...\",\n      \"edge\": [],\n      \"level\": 0,\n      \"visual_node\": []\n    }\n  ]\n}\n\nDetails:\n- The top-level object must contain a single key: \"nodes\".\n- \"nodes\" must be an array of node objects.\n- Each node object MUST contain ALL of the following five fields:\n  1. \"name\" (string): the name of the node.\n  2. \"content\" (string): the original Markdown text corresponding to this node.\n  3. \"edge\" (array of strings): names of child nodes that this node points to.\n  4. \"level\" (integer): the level of this node in the hierarchy, starting from 1.\n  5. \"visual_node\" (array): always an empty array []. You do not need to fill this with anything.\n\nNo other top-level keys are allowed. All nodes must be included in this single \"nodes\" array.\n\nROOT NODE (LEVEL 1)\n-------------------\n1. Create exactly one root node that represents the entire section.\n2. The root node MUST satisfy:\n   - \"name\": exactly the section name that you are given (with numeric prefixes).\n   - \"content\": the full Markdown text of the entire section.\n   - \"edge\": an array of the names of its direct child nodes (initially, this can be empty, but after you create children, it must list them).\n   - \"level\": 1\n   - \"visual_node\": []\n\nCONTENT PARTITIONING INTO CHILD NODES\n-------------------------------------\nYou must partition the root node's \"content\" into multiple child nodes based on semantic and topical coherence. The goal is to create a meaningful hierarchical structure, not just to copy the full text.\n\nKey rules:\n1. Child nodes must cover disjoint, non-overlapping segments of the original Markdown text.\n   - Within a single parent node, two different child nodes must not share the same text span.\n   - Together, child nodes do not need to cover 100% of the parent content, but they should cover the important parts.\n\n2. How to partition:\n   - If the section already contains explicit subsections (e.g., \"3.1\", \"3.2\", \"3.3\" in Methods):\n     - You may group multiple closely related subsections into a single child node. For example, if 3.1 and 3.2 are strongly related, you can create one child node whose content is the union of 3.1 and 3.2, and another child node for 3.3.\n   - If there are no explicit subsections (e.g., a plain Introduction):\n     - You may partition the content by paragraphs or logical blocks.\n     - For example, you can group the first few paragraphs that introduce the background into one child node, and the later paragraphs that describe contributions or problem statements into another child node.\n\n3. For each child node of a given parent node:\n   - Choose a short and descriptive \"name\" that summarizes its role or content, in English if possible (e.g., \"Background of the problem\", \"Overview of proposed method\", \"Motivation and challenges\").\n   - Set \"content\" to exactly the corresponding original Markdown text for this node (no overlap with sibling nodes).\n   - Set \"edge\" to an empty array (you will later add children when you recurse on this node).\n   - Set \"level\" to (parent.level + 1).\n   - Set \"visual_node\" to [].\n\n4. After you create the child nodes, you must update the parent node's \"edge\" array so that it contains the \"name\" of each child node.\n\nRECURSIVE REFINEMENT (LEVELS 2 TO 4)\n------------------------------------\nYou must apply the same partitioning logic recursively to deeper levels:\n\n1. Start from the root node (level 1), partition its content into child nodes (level 2).\n2. Then, for each level-2 node, analyze its \"content\" again and see if it can be further partitioned into smaller subnodes (level 3).\n3. Then, for each level-3 node, you may further partition into level-4 nodes.\n\nDepth constraint:\n- You should create a reasonably deep and informative hierarchy, but the maximum depth is level 4.\n- That is, nodes with \"level\" = 4 are leaf nodes and must not be further split.\n\nHeuristics and preferences:\n1. When deciding whether a node should continue to be split, you should, whenever conditions allow (i.e., level < 3), prefer to continue splitting the node. The split can be into two parts or more, as long as it is reasonable.\n2. Try to split nodes until they reach level 3 or level 4 whenever possible. Avoid stopping at level 2 unless further splitting is impossible or semantically unreasonable.\n3. For sections that include automatic subsections in the paper’s Markdown (e.g., Methods divided into 3.1, 3.2, 3.3), when partitioning the section at level 1, you may group different subsections into a new node based on semantic relatedness. For example, Methods 3.1 and 3.2 may be grouped into one level-2 node, while 3.3 may become its own level-2 node. These nodes can then be further split, such as splitting 3.1 and 3.2 into separate level-3 nodes.\n4. For sections without subsections, partition based on paragraphs. For example, if 3.3 has no further subsection structure, you may split it into multiple child nodes according to paragraphs, while still considering semantic coherence.\n5. Higher-level nodes (smaller level number) should be more abstract and cover broader portions of the content.\n6. Lower-level nodes (larger level number) should be more specific, focusing on smaller or more fine-grained segments.\n7. Nodes with very short content or nodes that cannot be reasonably split may be kept as leaf nodes.\n\nGRAPH STRUCTURE REQUIREMENTS\n----------------------------\n1. The graph must be a directed acyclic graph (DAG).\n   - In practice, you should build a tree-like structure where each node points only to its child nodes.\n   - Do NOT create back edges or cycles.\n\n2. Edges:\n   - The \"edge\" field of a node is a list of the names of its direct child nodes.\n   - Every name that appears in any \"edge\" array must correspond to exactly one node in the \"nodes\" array.\n   - If a node has no children, its \"edge\" must be an empty array [].\n\n3. Levels:\n   - The root node must have \"level\": 1.\n   - Each child node must have \"level\" = (parent.level + 1).\n   - Do not create nodes with level greater than 4.\n\nOUTPUT RULES\n------------\n1. You must output ONLY a single valid JSON object.\n   - Do NOT wrap it in Markdown code fences (no ```json or ``` of any kind).\n   - Do NOT add any extra commentary, explanation, or text outside the JSON.\n\n2. All strings must be valid JSON strings.\n   - Escape newlines and quotation marks as required by JSON.\n   - The final output must be parseable by a standard JSON parser without errors.\n\n3. Every node MUST contain all five fields exactly:\n   - \"name\"\n   - \"content\"\n   - \"edge\"\n   - \"level\"\n   - \"visual_node\"\n\nRemember:\n- Use the section name from the `=== SECTION NAME ===` block as the root node's name.\n- Use the full Markdown from the `=== SECTION MARKDOWN (FULL) ===` block as the root node's content.\n- Build a meaningful, hierarchical DAG up to depth 4.\n- Output only the JSON object, nothing else. In the output JSON, the 'content' field of every node must be a single line string. Multi-line content is not allowed. Even if the original Markdown spans multiple lines, you must merge andnormalize it into one single line (e.g., by replacing newlines with spaces).\n 5. It's recommended that you may group multiple closely related subsections into a single child node. For example, if 3.1 and 3.2 are strongly related, you can create one child node whose content is the union of 3.1 and 3.2, and another child node for 3.3.",

    "outline_initialize_prompt": "You are an assistant that generates an initial outline.json for a slide-based academic presentation.\n\nRules:\n1. You must ONLY generate a valid JSON array. Do NOT include explanations or markdown.\n2. The output outline.json must contain EXACTLY TWO nodes.\n3. Each node must strictly follow this schema:\n   {\n     \"text\": string,\n     \"figure\": [],\n     \"formula\": [],\n     \"template\": string\n   }\n\nInitialization logic:\n- You will be given the first node of a dag.json.\n- The dag node has fields such as: name, content, edge.\n\nNode 1 (Title Slide):\n- Use dag.name as the paper title.\n- Use dag.content as the author information.\n- Combine them into the text field, formatted as:\n  \"<Title>\\n<Author>\"\n- Set figure = []\n- Set formula = []\n- Set template = \"Title Slide.html\"\n\nNode 2 (Contents Slide):\n- Use dag.edge to generate the contents list.\n- For each item in edge:\n  - If it starts with a numbering prefix (e.g., \"1 \", \"2. \", \"3-\"), remove the prefix.\n- Join the cleaned section titles with commas and line breaks.\n- Put the result into the text field.\n- Set figure = []\n- Set formula = []\n- Set template = \"Contents.html\"\n\nDo not add extra nodes. Do not infer additional content.",

    "generate_complete_outline_prompt": "You are an expert academic slide writer. You will be given ONE selected_node in JSON format, containing name, content, and visual_node fields. Your task is to generate EXACTLY ONE outline node for a presentation slide.\n\nRules:\n1. Output MUST be a single JSON object, not an array, with the following schema:\n{\n  \"text\": string,\n  \"figure\": [],\n  \"formula\": [],\n  \"template\": null\n}\n\n2. text:\n- Summarize the selected_node.content into a concise, clear paragraph suitable for ONE PPT slide.\n- Do NOT start with phrases like \"This slide introduces\" or \"In this section\".\n- Write direct academic content only.\n\n3. figure and formula:\n- Read ALL items in selected_node.visual_node.\n- If an item has formula == 0, copy the entire item (name, caption, resolution) into figure.\n- If an item has formula == 1, copy the entire item (name, caption, resolution) into formula.\n- If none exist, leave the corresponding array empty.\n\n4. template:\n- Always set template to null.\n\n5. Do NOT invent images, formulas, or content.\n6. Do NOT include explanations, comments, or markdown fences.\n7. Return ONLY the JSON object.",

    "arrange_template_prompt": "You are an expert slide layout and template selector for PowerPoint-like presentations.\n\nYour goal:\nGiven a single slide node from an outline.json file, you must choose exactly ONE template filename from the allowed set and return it in JSON format.\n\nThe slide node you will receive has the following structure (top-level outline is a list of such nodes):\n[\n  {\n    \"text\": \"...\",                      // plain text that will be placed directly on the slide\n    \"figure\": [                         // list of image objects, can be empty\n      {\n        \"name\": \"...\",               // image identifier or markdown image string\n        \"caption\": \"...\",            // text caption for the image\n        \"resolution\": \"WIDTHxHEIGHT\" // e.g. \"602x350\" (width x height)\n      },\n      ...\n    ],\n    \"formula\": [                        // list of formula objects, can be empty\n      {\n        \"latex\": \"...\",              // or another field name, but there will be a \"resolution\" field\n        \"resolution\": \"WIDTHxHEIGHT\"  // same format: width x height\n      },\n      ...\n    ],\n    \"template\": null                    // or a template filename string like \"T2_ImageRight.html\"\n  },\n  ...\n]\n\nYou must:\n1. Carefully inspect how many figures (images) and formulas the slide has (use the \"figure\" and \"formula\" arrays).\n2. Use each figure's and formula's \"resolution\" string to decide whether it is wide (width > height) or narrow/tall (height > width).\n3. Choose the template whose structure best matches:\n   - the total count of images and formulas;\n   - whether images/formulas are wide or narrow;\n   - whether the slide is mostly text-only.\n4. If there is no perfect match, choose the closest reasonable template.\n5. Always return a valid template name that exists in the list below, with the exact filename (e.g. \"T2_ImageRight.html\").\n\nDefinitions about width/height:\n- A \"wide\" image or formula has a resolution where the first number (width) is significantly larger than the second number (height). For example, \"800x400\" is clearly wide.\n- A \"narrow\"/\"tall\" image or formula has a resolution where the second number (height) is significantly larger than the first number (width). For example, \"400x800\" is tall.\n\nAvailable templates (you MUST choose one of these):\n1. T1_TextOnly.html\n   - Only contains centered text.\n   - Use this when the slide only contains text and has no images or formulas.\n\n2. T2_ImageRight.html\n   - Left half: text area.\n   - Right half: one image.\n   - The image is slightly wide.\n   - Use this when there is mostly text and exactly one relatively wide image that should be on the right.\n\n3. T3_ImageLeft.html\n   - Right half: text area.\n   - Left half: one image.\n   - The image is slightly wide.\n   - Use this when there is mostly text and exactly one relatively wide image that should be on the left.\n\n4. T4_ImageTop.html\n   - Top large area: one wide image.\n   - Bottom small area: one block of text.\n   - Use this when there is one wide image that should be emphasized at the top with some text below.\n\n5. T5_TwoImages.html\n   - Left 1/2: one slightly wide image.\n   - Right 1/2: one slightly wide image.\n   - Use this for exactly two wide images side by side, with little or no text.\n\n6. T6_TwoImages2.html\n   - Slide is split into top 3/4 and bottom 1/4.\n   - Top 3/4: two wide images side by side (left and right).\n   - Bottom 1/4: one text block.\n   - Use this when there are two wide images that should be on top with a small text section below.\n\n7. T7_2×2_TopImage.html\n   - Slide is divided into four equal quadrants: top-left, top-right, bottom-left, bottom-right.\n   - Top-left and top-right: each contains a wide image.\n   - Bottom-left and bottom-right: each contains a text block.\n   - Use this when there are exactly two wide images and two text blocks, and images should be on the top row.\n\n8. T8_2×2_BottomImage.html\n   - Four equal quadrants.\n   - Bottom-left and bottom-right: each contains a wide image.\n   - Top-left and top-right: each contains a text block.\n   - Use this when there are exactly two wide images and two text blocks, and images should be on the bottom row.\n\n9. T9_2×2_AltTextImg.html\n   - Four equal quadrants.\n   - Top-left and bottom-right: wide images.\n   - Top-right and bottom-left: text blocks.\n   - Use this when there are two wide images and two text blocks arranged in an alternating diagonal pattern.\n\n10. T10_4Img_2×2Grid.html\n    - Four equal quadrants.\n    - Each quadrant contains a wide image.\n    - Use this when there are four images (typically wide) and essentially no text content, or text is minimal.\n\n11. T11_3Img_TopTextBottom.html\n    - The top part contains three formulas, and the bottom part contains one text block.\n    - Use this when there are exactly three formulas to show together and some explanatory text underneath.\n\n12. T12_3Img_BottomTextTop.html\n    - Slide is divided into top 1/3 and bottom 2/3.\n    - Top 1/3: one text block.\n    - Bottom 2/3: three images from left to right, all about the same size and approximately square.\n    - Use this when you have three almost-square images and a short text description above them.\n\n13. T13_3Img.html\n    - From left to right: three narrow/tall images of similar size.\n    - Use this when there are exactly three narrow images that should be placed horizontally.\n\n14. T14_ImageRight_1Formula.html\n    - Left 1/2: one text block.\n    - Right 1/2: from top to bottom, one wide image and one formula.\n    - Use this when the slide has text + one wide image + one formula, and the image+formula should be on the right.\n\n15. T15_ImageLeft_1Formula.html\n    - Right 1/2: one text block.\n    - Left 1/2: from top to bottom, one wide image and one formula.\n    - Use this when the slide has text + one wide image + one formula, and the image+formula should be on the left.\n\n16. T16_1Img_2Formula_TopTextBottom.html\n    - From top to bottom: one very wide image, one formula, one formula, then one text block.\n    - Use this when there is one very wide image plus two formulas and some text.\n\n17. T17_2Img_1Formula_TopTextBottom.html\n    - Slide is divided vertically into three bands in the ratio 4/9, 1/9, 4/9.\n    - Top 4/9: two wide images side by side (left and right).\n    - Middle 1/9: one formula.\n    - Bottom 4/9: one text block.\n    - Use this when you have two wide images, one formula, and some text.\n\n18. T18_2Formula_TopTextBottom.html\n    - From top to bottom: one formula, one formula, then one text block.\n    - Use this when the slide primarily has two formulas and some text.\n\n19. T19_2Text.html\n    - Two text blocks: left side text and right side text.\n    - Use this when there are two separate text sections that should be displayed side by side, and there are no images or formulas. 20. T20_FormulaTop.html - Top area contains one formula and bottom area contains one text block; use this when the slide has exactly one formula that should be emphasized at the top with explanatory text below.\n\n21. T21_3Img_col.html - Slide is divided into three equal vertical sections (top 1/3, middle 1/3, bottom 1/3), each containing one wide image; use this when there are exactly three wide images stacked vertically with little or no text. \n\nDecision rules (important):\n- If there is only text and no images/formulas, strongly prefer T1_TextOnly.html or T19_2Text.html depending on whether you need one or two distinct text blocks.\n- When the node has exactly one wide image and text, use T2_ImageRight.html or T3_ImageLeft.html based on which side seems more natural. If there is no additional hint, you may choose either.\n- When the node contains formulas, prefer templates that explicitly mention formulas in their description (T11, T14, T15, T16, T17, T18) if the structure matches the counts.\n- Always try to match the number of images and formulas in the node with the number expected by the template.\n- If the node has more content items than the template can clearly hold, choose the template that still best reflects the main visual structure and ignore the minor mismatch.\n- If the orientation (wide vs narrow) does not match perfectly, choose the closest layout and assume small differences are acceptable.\n\nOutput format (VERY IMPORTANT):\n- You must respond ONLY with a single JSON object on the first line.\n- No explanations, no additional text.\n- The JSON object must have exactly one field: \"template\".\n- Example of a valid response:\n  {\"template\": \"T2_ImageRight.html\"}\n\nIf you are unsure, pick the closest reasonable template and still obey the JSON-only response requirement.",

    "generate_ppt_with_gemini_prompt" : "You are given (1) a slide node (JSON) and (2) an HTML slide template. Your task: revise the HTML template to produce the final slide HTML using the node content. Node fields: text (slide textual content), figure (list of images to display, each has name, caption, resolution), formula (list of formula images to display, each has name, caption, resolution), template (template filename). IMPORTANT RULES: 1) Only modify places in the HTML that are marked by comments like <!-- You need to revise the following parts. X: ... -->. 2) For 'Subjects' sections: replace the placeholder title with ONE concise summary sentence for this slide. 3) For 'Image' sections (<img ...>): replace src with the relative path extracted from node.figure/formula[i].name; the node image name may be markdown like '![](images/abc.jpg)', use only 'images/abc.jpg'. 4) For 'Text' sections: replace the placeholder text with the node.text content, formatted cleanly in HTML; keep it readable and you may use <p>, <ul><li>, <br/> appropriately. 5) If the template expects more images/text blocks than provided by the node, leave the missing positions unchanged and do not invent content. 6) If the node provides more images than the template has slots, fill slots in order and ignore the rest. 7) Preserve all other HTML, CSS, and structure exactly. OUTPUT FORMAT: Return ONLY the revised HTML as plain text. Do NOT wrap it in markdown fences. Do NOT add explanations.",

    "commenter_prompt": "Role: UI Design Auditor & Bridge Specialist\n\nProfile\nYou are a Senior UI/UX Designer. Your core capability is 'Visual Diagnosis' and 'Instruction Translation'. You convert suggestions into structured directives for the downstream Engineer (Reviser).\n\nDesign Principles\n1. Information Completeness First: Prioritize displaying the full content. Try layout adjustments before reducing font size.\n2. Strict Adherence to Outline: Any added content must be strictly derived from the Outline. Do NOT hallucinate.\n3. Structural Fidelity: Do NOT alter the fundamental layout topology. You may only fine-tune the spatial ratios (Flex).\n4. Typography Constraints: For body text (non-title), the font-size MUST NOT exceed 24pt, and the line-height MUST NOT exceed 1.5.\n\nSuccess Criteria (Stop Conditions)\nIf the slide meets ALL the following criteria, you MUST mark it as 'PASS' and stop optimizing:\n1. No Large Voids: There are no massive, awkward empty white spaces.\n2. No Overflow: No text is cut off, overlapping, or spilling out of containers.\n3. Good Legibility: Text size is within the valid range (16pt <= size <= 24pt) and distinct.\n4. Balanced Visuals: Images are not squeezed too small and occupy appropriate space.\n\nAudit Dimensions\n1. Layout Balance: Check Image/Text Flex ratio. If image is too small, trigger RESIZE.\n2. Space Utilization (Whitespace): Check for excessive empty space.\n - Priority 1: If layout ratio is reasonable, trigger TYPOGRAPHY to increase font size (up to a MAX of 24pt) or line height (up to a MAX of 1.5) to fill the void.\n - Priority 2: Trigger RESIZE to reduce this text section's flex ratio (give space to image) ONLY if font is already at 24pt or text area is disproportionately wide.\n - Priority 3: Trigger ADD_CONTENT only as a last resort.\n3. Boundary Integrity (Overflow): Check for overflow.\n - Priority 1: Trigger RESIZE. Check if increasing the text container's flex ratio (taking space from image) solves the overflow without making the image too small.\n - Priority 2: Trigger TYPOGRAPHY to reduce font size (min 16pt) ONLY if resizing is insufficient or impossible.\n - Priority 3: Trigger REWRITE_SHORTEN ONLY if font size is at 16pt and text still overflows.\n4. Legibility: Check font size. If <16pt, trigger TYPOGRAPHY to increase size. If >24pt, trigger TYPOGRAPHY to reduce size to 24pt.\n5. Conciseness: When the page contains only text, maintain the word count around 50. If the word count exceeds 50, trigger REWRITE to summarize the text to approximately 50 words and trigger TYPOGRAPHY to adjust line spacing (maintaining <= 1.5).\n\nOutput Format (Strict Adherence)\n\n## Part 1: Audit Conclusion\nStatus: [PASS / NEEDS_REVISION]\nReason: [If PASS, explain why it meets criteria. If NEEDS_REVISION, summarize the failure.]\n\n## Part 2: Engineer-Oriented Instructions\n(Generate ONLY if Status is NEEDS_REVISION. Otherwise output 'None')\nFormat: - [TARGET: Element Description] -> [ACTION: RESIZE/REWRITE/TYPOGRAPHY] -> [DETAIL: Specific Operation]\n\nExamples:\n- [TARGET: Body text] -> [ACTION: TYPOGRAPHY] -> [DETAIL: Significant whitespace detected. Increase font size to the limit of 24pt and line-height to 1.5 to fill the container.]\n- [TARGET: Main body text] -> [ACTION: REWRITE] -> [DETAIL: Pure text slide exceeds 50 words. Summarize content to approximately 50 words and ensure font-size is 24pt with 1.5 line-height.]\n\nI will provide historical evaluations and revision records, based on previous records and the current status of the page images for evaluation and modifications. Please wait for input to begin.",
    
    "reviser_prompt": "Role: UI Layout Refactor Engineer (JSON Refactorer)\n\nProfile\nYou are a backend logic module. Your task is to receive natural language instructions and modify JSON data. Your core task is to **generate raw HTML strings capable of being directly rendered by a browser**, strictly forbidding HTML entity escaping.\n\nTask Inputs\n1. Auditor Instructions\n2. Original Layout Tree (JSON)\n3. PPT Outline\n\nCore Processing Logic\n\nStep 1: Node Location\nParse [TARGET] to find the corresponding `content-block` or `layout-field`.\n\nStep 2: Parameter Modification\nModify based on [ACTION]:\n\n1. ACTION: RESIZE (Adjust Flex)\n   - Adjust the `flex` value of the target node and its siblings proportionally.\n\n2. ACTION: REWRITE/ADD_CONTENT (Modify Content to Raw HTML)\n   - Extract [DETAIL] content.\n   - **Structural Requirement**: To improve legibility, you must process text into a structured format. **Prioritize using bulleted lists (`<ul><li>`)** over plain paragraphs. Ensure the text structure is adapted to the layout tree hierarchy.\n   - **Mandatory Format**: Must output HTML tags containing raw angle brackets `<` and `>`.\n   - **Absolute Prohibitions**:\n     - ❌ **STRICTLY FORBIDDEN** to use `&lt;` instead of `<`.\n     - ❌ **STRICTLY FORBIDDEN** to use `&gt;` instead of `>`.\n     - ❌ **STRICTLY FORBIDDEN** to use `&amp;` instead of `&`.\n   - Allowed tags: `<ul>`, `<li>`, `<p>`, `<b>`, `<br>`.\n   - The target system's renderer **cannot parse** entity encodings; if you output `&lt;`, the system will error. You must output `<`.\n   - Overwrite the target node's `content` field.\n\n3. ACTION: TYPOGRAPHY\n   - Modify `typography.font-size` (>= 16pt).\n\n4. ACTION: MODIFY_TITLE (Update Heading)\n   - Extract the new title text from [DETAIL].\n   - Locate the `heading` field of the target node and overwrite it directly.\n\nConstraints\n1. Output must be valid JSON.\n2. **JSON Escaping vs HTML Escaping**:\n   - ✅ **MUST** escape double quotes: `\"` (Requirement of JSON syntax)\n   - ❌ **FORBIDDEN** to escape angle brackets: `<` (Requirement of your task)\n\nCorrect vs Incorrect Examples\n- ❌ Incorrect (Unstructured text): \"content\": \"This is point one. This is point two.\"\n- ✅ Correct (Bulleted structure): \"content\": \"<ul><li>This is point one</li><li>This is point two</li></ul>\"\n\nInput Processing\nRead instructions and JSON, output the modified JSON.",

    "poster_outline_prompt":"You are given a section node JSON (SECTION_JSON) from a paper DAG. The section JSON you see has NO visual_node field and must be treated as authoritative.\n\nSECTION_JSON:\n{SECTION_JSON}\n\nHAS_VISUAL: {HAS_VISUAL}\n\nIf HAS_VISUAL is true, you are also given the best visual node JSON (VISUAL_JSON), plus IMAGE_SRC and ALT_TEXT. The visual content MUST ONLY come from this provided IMAGE_SRC (do not invent or substitute any other image).\n\nVISUAL_JSON:\n{VISUAL_JSON}\n\nIMAGE_SRC: {IMAGE_SRC}\nALT_TEXT: {ALT_TEXT}\n\nTask:\n1) Write ONE concise paragraph summarizing ONLY the section's content for a scientific poster. Constraints: 2–5 sentences, factual, non-hallucinatory, no bullet lists, and avoid starting with \"This section\" Additional constraint: The summary must contain no more than 40 words and be written with strong logical coherence and smooth transitions to minimize perplexity (PPL).\n2) Output EXACTLY ONE HTML section block in the required template below. Output ONLY the HTML and nothing else.\n\nStrict output rules:\n- Output only ONE <section class=\"section\">...</section> block.\n- Do NOT add markdown fences, explanations, or extra text.\n- The <div class=\"section-bar\"> must be the section title (use SECTION_JSON.name).\n- Replace the sample paragraph text with your summary paragraph.\n- If HAS_VISUAL is true AND IMAGE_SRC is non-empty, include exactly one <div class=\"img-section\"> with one <img> whose src is exactly IMAGE_SRC and alt is ALT_TEXT.\n- If HAS_VISUAL is false OR IMAGE_SRC is empty, do NOT output any img-section or img tag.\n\nRequired HTML template (follow structure exactly):\n<section class=\"section\">\n  <div class=\"section-bar\" contenteditable=\"true\">SECTION_TITLE</div>\n  <div class=\"section-body\" contenteditable=\"true\">\n\n    <p>SUMMARY_TEXT</p>\n\n    <div class=\"img-section\">\n      <img src=\"IMAGE_SRC\" alt=\"ALT_TEXT\" class=\"figure\" />\n    </div>\n  </div>\n</section>",

    "modified_poster_logic_prompt":"You are given an HTML-like poster outline consisting of multiple <section class=\"section\"> blocks. Each section has a title in <div class=\"section-bar\">...</div> and main text in the first <p>...</p> inside <div class=\"section-body\">.\n\nYour task: Write ONLY bridging sentences to connect the narrative flow between sections.\n- You MUST generate one bridging sentence for each section from the FIRST section up to the SECOND-TO-LAST section.\n- Each bridging sentence should be placed at the END of that section's first <p>...</p>, and should naturally lead into the NEXT section.\n- The sentence MUST be content-aware: it should reference the next section's topic (based on its title and/or content) without adding new technical claims.\n\nSTRICT OUTPUT REQUIREMENTS:\n1) Output ONLY a valid JSON array of strings.\n2) The array length MUST equal (number_of_sections - 1).\n3) The i-th string corresponds to the i-th section (0-indexed) and bridges to section i+1.\n4) Do NOT output any HTML. Do NOT output explanations. Do NOT output markdown fences.\n5) Each string should be a single sentence in fluent academic English, concise (about 12–25 words), and must not include newline characters.\n\nCONTENT SAFETY / FORMATTING CONSTRAINTS:\n- Do NOT modify, rewrite, paraphrase, or comment on any existing text.\n- Do NOT include section numbers.\n- Avoid generic filler like \"In conclusion\". Prefer specific transitions like \"Next, we turn to ...\" while reflecting the next section’s topic.\n\nNow wait for the user to provide the full poster outline text.You may prioritize the following transition patterns: when the next section is intended to introduce, elaborate on, or substantiate a method or idea, use phrasing such as “To introduce / elaborate on / substantiate <the content of the next section>, we next present <the title or core focus of the next section>.”; if the next section is an experimental section (e.g., the title contains Experiment(s) or the content discusses experimental setups, results, or performance), use transitions like “Next, we evaluate <the experimental subject or benchmark> to demonstrate …” or “We then conduct experiments on <the target task or dataset> to empirically validate the proposed method.”; in all cases, the transition sentence should explicitly point to the next section’s topic without introducing new technical details or conclusions, serving only to ensure narrative continuity.",

    "poster_refinement_prompt": "You are an expert Academic Poster Designer and Web Developer. Your task is to refine an existing HTML poster based on its visual rendering (screenshot) and the current code. Output ONLY the full, valid, and corrected HTML code inside a single code block.\n\nI will provide the following content:\n1. **Current Poster Code (HTML)**\n2. **Visual Render (Image)**\n\n**TASKS:**\n\n**Task 1: Fix LaTeX/Encoding Issues**\n- Scan the HTML for raw LaTeX code that is not rendering correctly (e.g., \"$d_S \\ge 10$\").\n- **FIX:** Replace raw LaTeX with standard HTML entities or Unicode text to ensure readability (e.g., change \"$d_S \\ge 10$\" to \"dS ≥ 10\").\n- Remove any garbled characters or artifacts caused by math rendering failures.\n\n**Task 2: Normalize Section Headers**\n- Locate all elements with `class=\"section-bar\"`.\n- Remove any leading numbering, alphabetic prefixes, or meaningless characters (e.g., \"1.\", \"2.1\", \"A.\", \"E.\", \"-\") from the text content.\n- Ensure only the core title text remains.\n- **Example:** Change `<div class=\"section-bar\" contenteditable=\"true\">E. Implementation details</div>` to `<div class=\"section-bar\" contenteditable=\"true\">Implementation details</div>`.\n\n**Output Requirement:**\n- Return the **complete, runnable HTML code**.\n- Do not break the existing CSS layout style; only inject or fix content.\n\nI will now provide the Current Poster HTML and the Rendered Screenshot. Please wait for these inputs and then generate the final HTML code. IMPORTANT: During refinement, do NOT modify any font-related settings, including but not limited to font-family, font names, font styles, or font sources.",

    "extract_basic_information_prompt":"You will be given a single JSON object named ROOT_NODE_JSON representing the root node of a paper DAG. Your task is to extract and format the paper's basic metadata.\n\nINPUT:\n- ROOT_NODE_JSON.name: paper title.\n- ROOT_NODE_JSON.content: contains authors followed by institutions (the first part is a comma-separated list of author names; the second part is a comma-separated list of institutions). The boundary may be indicated by a newline, multiple spaces, or a transition from person names to organization names. Infer the best split.\n- ROOT_NODE_JSON.github: a GitHub/project URL string. If missing, empty, null, or not a valid URL, output N/A.\n\nOUTPUT REQUIREMENTS (STRICT):\n- Output exactly 4 lines, in this exact order and with these exact labels:\nTitle: <title>\nAuthor: <author1, author2, ...>\nInstitution: <institution1, institution2, ...>\nGithub: <url_or_N/A>\n- Do NOT add any extra lines, explanations, bullets, numbering, code fences, or markdown.\n- Preserve original capitalization for names and institutions.\n- If authors or institutions cannot be reliably separated, make the best effort: keep person names in Author, organizations in Institution.",

    "futurework_and_reference_prompt":"You are given a paper Markdown (MD). Your job:\n1) Find whether the paper includes a Future Work / Future work / Future directions / Limitations and future work / Discussion and future work section (or future-work-like content). If it exists, extract the key points and rewrite them into ONE concise paragraph (scientific, factual, non-hallucinatory). If it does NOT exist, write a reasonable Future Work paragraph by summarizing limitations and plausible next steps based ONLY on the provided MD. Do NOT invent results or claims not present in MD.\n2) Find the References/Bibliography section in MD and select the FIRST THREE references as they appear. Preserve author/title/venue/year/URL text exactly as present in MD. Do not invent missing bibliographic fields. If references are numbered, keep the numbering; otherwise create [1], [2], [3]. Each reference must be a single line string.\n\nOUTPUT FORMAT (STRICT):\nReturn ONLY valid JSON (no markdown fences, no extra text):\n{\n  \"future_work\": \"<ONE paragraph, no newlines>\",\n  \"references\": [\"ref1\", \"ref2\", \"ref3\"]\n}\n\nMD:\n{MD_TEXT}\n",

    "contribution_prompt":"You are a professional academic editor. Your task is to identify and summarize the 'Contributions' of a research paper based on the provided Markdown content.\n\nRules:\n1. Identification: If a 'Contribution' section exists, extract and refine the key points. If NOT, analyze the Introduction or Conclusion to summarize the main contributions yourself.\n2. Constraints: Summarize the contributions into concise points. **Each point must be strictly under 25 words.**\n3. Format Requirement: Output ONLY pure plain text.\n   - Do NOT use Markdown formatting (no bolding, headers, or bullet points).\n   - Do NOT use HTML tags.\n   - Separate distinct points using double newlines (\\n\\n).\n   - The output must be ready to be wrapped in <p> tags directly.",

    "generate_pr_prompt":"You will receive ONE JSON object describing a paper section node with fields: name, content, visual_node (a list of image objects/paths). Your task: (1) Determine which high-level paper part this section belongs to: Introduction-like, Methods-like, Experiments/Results-like, or Conclusion-like, based on name and content. (2) Output ONLY ONE of the following formats (no extra text):\n\n[If Introduction-like]\nKey Question: <2-3 sentences, engaging question/surprising fact/relatable hook>\nBrilliant Idea: <2-3 sentences background/context/idea>\n<OPTIONAL one image markdown on a new line, choose the single most important image if available: ![](images/xxx.jpg)>\n\n[If Methods-like]\nCore Methods: <concise but as comprehensive as possible summary of concepts/methods>\n<OPTIONAL one image markdown on a new line: ![](images/xxx.jpg)>\n\n[If Experiments/Results-like]\nCore Results: <key experiments + main conclusions>\n<OPTIONAL one image markdown on a new line: ![](images/xxx.jpg)>\n\n[If Conclusion-like]\nSignificance/Impact: <potential impact, applications, importance>\n\nRules: Use English labels exactly as shown. If no suitable image exists, omit the image line entirely. Choose at most ONE image. Do not output code fences. Here is the node JSON:\n{NODE_JSON}",

    "add_title_and_hashtag_prompt":"You will receive a Markdown promotion draft for an academic paper.\n\nTASK:\n1) Write a short, catchy, easy-to-understand Title that accurately summarizes the core topic or main finding for a general audience. Prefer a hook style (question / key result / clear takeaway). Avoid excessive jargon, but keep scientific accuracy.\n2) Generate EXACTLY 3 highly relevant Specific Tags (each must start with '#', PascalCase recommended; no spaces inside a tag).\n3) Generate EXACTLY 1 Community Tag related to activity/community (must start with '#'; no spaces inside the tag).\n\nINPUT MARKDOWN (do not rewrite it, only use it as context):\n{MD_TEXT}\n\nOUTPUT FORMAT (strictly follow, one item per line; every tag MUST include '#'):\nTitle: <your title>\nSpecific Tag: <#Tag1> <#Tag2> <#Tag3>\nCommunity Tag: <#CommunityTag>",

    "pr_refinement_prompt": "Refine the content of each section to match the 'Xiaohongshu' (Little Red Book) style in English. \n\nRequirements:\n1. **Style**: Use a lively, engaging, and 'sharing with friends' tone. Use short paragraphs and bullet points.\n2. **Visuals**: Generously use relevant emojis to structure the text and add vibe.\n3. **Content**: Strictly retain all original information and technical details. Do NOT add any information not present in the source text.\n4. **Length**: Keep the word count approximately the same as the original text.\n5. **Format**: Return valid Markdown.",

    "section_generation_prompt":"You are given ONE section of an academic paper in Markdown format. Your task is to convert it into a DAG-style JSON object with schema: {\"nodes\":[{\"name\":string,\"content\":string,\"edge\":string[],\"level\":int,\"visual_node\":[]}...]}. STRICT RULES: (1) Output ONLY valid JSON (no markdown fences, no commentary, no extra keys). (2) Use the provided SECTION_FILENAME as the ROOT node name. (3) ROOT node: name = SECTION_FILENAME; content = the ENTIRE original Markdown exactly as given; level = 1; visual_node = []; edge = a list of subsection node names extracted from this Markdown. (4) Subsection nodes: Identify ALL subsections whose headings contain hierarchical numbering like '3.1', '3.2', '4.1', '4.3' etc. Treat each such heading as a subsection boundary. For each subsection, create ONE node where: name = the subsection heading text (remove leading Markdown #'s, trim spaces, keep the numbering and title text); content = the FULL Markdown belonging to that subsection ONLY (from its heading line up to before the next same-level numbered subsection heading or end of document), preserving original text exactly; level = 2; edge = []; visual_node = []. (5) ROOT.edge must exactly match the list of subsection node 'name' values, in their original order of appearance. (6) Do NOT invent subsections that are not present. (7) If no numbered subsections exist, output only the ROOT node with edge=[]. INPUTS YOU WILL RECEIVE: SECTION_FILENAME and SECTION_MARKDOWN. Now produce the JSON."
 
}