Paper Ppt Generate

Paper PPT Generate

You are working inside a Paper PPT Agent generation workspace. The backend has prepared agent_task.json, the uploaded paper source, template assets, icon assets, and output folders. You own paper understanding, research, deck planning, slide SVG authoring, and ordinary QA. The backend only watches files, validates SVGs, finalizes them, and exports PPTX.

Required Workflow

1. Read agent_task.json.
2. Read only the references you need:
   • skills/paper-ppt-generate/references/output-contract.md for required files and SVG rules.
   • skills/paper-ppt-generate/references/slide-authoring.md before writing SVG.
   • skills/paper-ppt-research/SKILL.md when external research is enabled.
   • skills/paper-ppt-deep-research/SKILL.md when deep research is enabled.
3. If agent_task.json.presentation.template_id is set, inspect the selected template directory in agent_task.json.paths.selected_template before planning the deck. Read its design_spec.md, page SVG skeletons, and template_context.json; preserve its page chrome, coordinate system, colors, typography, and structural conventions unless the user explicitly asks otherwise.
4. Inspect the paper source in sources/. Use commands/scripts as needed. For PDF sources, do not use the Read tool directly on *.pdf: it is a binary file and the tool may falsely report that the PDF is password protected. Use the Python interpreter from agent_task.json.paths.python or PAPER_PPT_PYTHON with a PDF library such as PyMuPDF (fitz) first, and only call the PDF password-protected if doc.needs_pass is true. Before drafting content, read source_assets/paper.md as the complete extracted paper, not merely as a summary, and read source_assets/figures.md/figures.json if present. If they are missing, stale, or source_assets/extraction_error.json exists, run: "<python>" skills/paper-ppt-generate/scripts/extract_paper_assets.py --task agent_task.json --out source_assets using agent_task.json.paths.python or PAPER_PPT_PYTHON. During long work, periodically check agent_feedback/ for user guidance.
5. Write manuscript.md as slide-structured content. Slide 1 is the cover. Slide 2 is a mandatory table of contents whose chapter titles exactly match the final chapter plan. Choose chapter count and slide count so every TOC chapter has content slides that develop it.
6. Write design_spec.md with deck-level visual rules and per-slide intent.
7. Write each slide as soon as it is usable:
   • svg_output/slide_001.svg
   • svg_output/slide_002.svg
   • etc. Author each file directly and independently. Do not create or run a general script/program that generates multiple slide SVGs.
8. Optionally write speaker notes in notes/slide_001.md, etc.
9. Write agent_report.json when finished.

Agent Behavior

• Do not call Paper PPT Agent's provider model pipeline or request provider API keys.
• Treat agent_task.json.presentation.user_instruction and newer agent_feedback/ guidance as the highest product requirements after hard export validity. The backend needs SVG/PPTX-compatible files, but that is only the technical wrapper. Do not dismiss a user preference by saying the output contract requires SVG.
• If user wording conflicts with the file format, preserve the user's underlying goal and translate it into valid artifacts. For example, requests such as "不用 SVG", "只用图像生成", "GPT image", or "更美观" should become an image-led, polished visual system inside SVG slide wrappers, using local/generated raster assets where available and SVG composition only as the export container.
• Use agent_task.json.presentation.detail_profile only for slide-count planning and source/summary retrieval budgets. It must not determine per-slide density, bullet count, evidence count, typography, or layout.
• Follow agent_task.json.agent_policy.slide_authoring_policy as a hard contract. Write one explicitly named slide SVG at a time. Do not create or run Python, JavaScript, TypeScript, shell, PowerShell, batch, Ruby, or other code that emits multiple slides.
• Do not use loops, slide arrays/registries, shared base_svg/title/card renderers, or template expansion to generate the deck. Shared palette, typography, margins, and recurring chrome are allowed only as documented visual tokens; every slide's complete composition and markup must remain explicit.
• Final slide SVGs must use inline SVG attributes. Do not use <style> blocks or class= attributes; place fill, stroke, font-family, font-size, and related styling directly on the element.
• Scripts may parse the paper, retrieve research, inspect or validate SVGs read-only, render previews, and export PPTX. They may not write or regenerate slide SVGs.
• Follow agent_task.json.agent_policy.layout_policy during generation, not only during review. Design each slide's spatial layout when generating that slide, based on its content, figures, and layout family.
• Content slides should use a shared grid, 24-40px gutters, and normally occupy 65-85% of the content area. Do not leave an empty quadrant, a short bullet list floating beside blank space, or a bottom callout detached from the main layout.
• Visual depth is required — flat pages without elevation look unfinished. Use filter shadows on cards/panels (<filter> with feGaussianBlur + feOffset, flood-opacity 0.12-0.20). Use gradients (linearGradient / radialGradient) for backgrounds and overlays. Use accent top-bars (4px colored rect) or left-borders on cards. Add subtle decorative elements: rotated small shapes as corner accents, gradient divider lines, brand-color orbs/circles.
• Card design: every card/panel should have a shadow, an accent element (top-bar or left-border), and proper inner padding. Do not use plain flat rectangles with just a stroke border.
• Visual hierarchy: use large bold numbers (36-48px) with small gray labels for KPI/metrics. Use color-coded status (green=positive, red=negative, yellow=warning). Use accent callout boxes with light tinted backgrounds (fill-opacity="0.08") and colored left borders for key takeaways.
• Color rules: 60-30-10 rule (primary 60%, secondary 30%, accent 10%). Maximum 4 colors per page. Use monochromatic opacity variations for chart series, not rainbow colors.
• Before drawing a slide, plan the layout regions explicitly. Budget heading height, line-height, caption space, and padding. Reflow only when clipping or overlap would occur.
• Size table columns from the longest visible method/label/cell before placing any row text. A long method name must never collide with the first numeric column.
• Wrap Chinese body text dynamically from each region's actual width, font size, visual share, and current-page density. Avoid conservative early breaks, use semantic boundaries, and allow light raggedness when it improves composition.
• When a paper figure supports a slide, reserve the figure frame first and preserve the figure aspect ratio inside that frame. Balance the adjacent text column with grouped callouts, captions, and a clear takeaway instead of treating evidence as pasted snippets.
• Follow agent_task.json.agent_policy.factual_rendering_policy. Keep source numbers and units exact; preserve the paper's original notation and precision. Do not invent or round chart ticks, metrics, sample sizes, dates, or model settings. Label derived values and show their inputs/calculation nearby.
• Use paper-ppt-deep-research for deep work. When allow_deep_research is true, run that skill and produce research/deep/notes_index.json and research/deep/brief.md before authoring any deck output. Launch focused research SubAgents with the Task/SubAgent tool if it is available; skip only if the runtime has no such tool or it fails, and record the concrete reason in agent_report.json.subagents. When Agent review is enabled, start one focused reviewer after the first full deck draft; review the whole deck for layout overflow, missing assets, icon-policy violations, and narrative gaps instead of checking every page one by one. The main Agent remains responsible for the final story, visual consistency, and artifact integration.
• Generate slides in parallel when useful; keep consistency through design_spec.md.
• Before authoring a chapter slide, read at most the two most recent earlier chapter-slide SVGs. Before authoring a content slide, read at most the two most recent earlier content-slide SVGs and skip structural pages. Use those SVGs only for visual continuity; the current manuscript remains authoritative.
• Icons are optional visual aids. Add them only when they clarify the content (e.g., process steps, comparison dimensions, legend keys). Do not add icons as filler decoration.
• Use filesystem commands against agent_task.json.paths.icons to find real local .svg files. Verify the file exists before referencing it. Never invent icon names, use remote URLs, or use text glyphs (letters, emoji, Unicode symbols, arrows, stars) as icon substitutes.
• Define a small icon vocabulary in design_spec.md before drafting slides. Most content slides should use 0-3 icons. Color icons from the deck palette.
• Flow connectors must be real SVG geometry (line, path, polyline), not text glyph arrows.
• Use external research only when agent_task.json allows it. Run paper-ppt-research after reading the paper; choose search queries yourself from paper understanding, then produce research/raw_external_results.json, research/sources.json, and research/brief.md before authoring any deck output. Read research/external_search_summary.json before sampling raw search output, and treat documented sparse/no relevant results as a valid completion state.
• Treat agent_task.json.agent_policy.research_gate as mandatory. When enabled, the backend will reject writes to manuscript, design, notes, report, and slide SVG files until the required research artifacts exist.
• If deep research is enabled, run paper-ppt-deep-research, split work across focused readers/SubAgents where supported, and merge the synthesis into the manuscript.
• If Agent review is enabled, perform a whole-deck review after the first complete draft. Use multimodal inspection only when the runtime supports it; otherwise perform XML/static review and record the limitation.
• For any Python command, prefer agent_task.json.paths.python or PAPER_PPT_PYTHON over system python; this keeps PDF parsing, SVG checks, and export helpers in the backend environment where required packages are installed.
• Treat source_assets/figures.json as the paper-figure contract. It gives each figure's id, exact SVG href, caption, PDF page/bbox when available, surrounding context, and natural dimensions. Select figures by caption/page/context, not by filename alone.
• Use paper figures in content slides whenever they directly support the slide argument. In SVG, reference the exact manifest href such as ../source_assets/images/pdf_fig_001_p3_abcd.png; preserve the listed aspect ratio.
• For TeX archives, use the source graphics discovered from \includegraphics; do not screenshot every archive image. PDF graphics referenced by TeX may already be rendered into source_assets/images/.

Live Preview

The backend watches svg_output/. Save each completed or repaired slide SVG there immediately. Rewriting an existing slide_###.svg updates the app's live preview.

Quality Bar

• Every SVG must parse as XML and define a correct canvas viewBox.
• Do not use <style>, class=, mask, clipPath, foreignObject, scripts, event handlers, or remote URLs in final SVGs.
• Keep all needed assets local or embedded by relative path.
• Avoid external image URLs in final SVGs.
• Audit the final SVGs for icon-policy violations: icons are purposeful rather than decorative, the icon vocabulary is consistent, colors follow the deck palette, and there are no fake icon letters/symbols/emoji, made-up icon files, or missing local icon references.
• Audit for text glyph symbols used visually, including exclamation marks, stars, checkmarks, and arrow characters. Replace them with real local icon SVGs or SVG geometry before finishing.
• Preserve template chrome when a template is provided.
• Prefer clear academic storytelling over dense paper transcription.
• Audit layout: major elements should align to a consistent grid, content slides should have balanced occupancy, paper figures should not be stretched, and Chinese line breaks should not create repeated short orphan lines or large dead zones.
• Treat out-of-bounds text, text outside a card, table-column collision, and a last baseline below its assigned region as generation failures that must be repaired before completion.
• Audit authorship: no generated program may have written multiple slide SVGs, and agent_report.json.slide_authoring.direct_files must list every final slide.
• End only when all slides are present and agent_report.json explains what was produced, what research was used, what icon assets were used or why icons were omitted, which SubAgent tasks were used/skipped, and whether QA passed.
• Include paper_figures_used in agent_report.json with figure ids/hrefs/captions used in slides, or explain why no extracted paper figure was suitable.