Paper to Slides#

Two-phase pipeline: Phase 1 deep-reads an academic paper into a structured dual-mode report, Phase 2 transforms the report into a polished, zero-dependency HTML presentation.

Dependencies#

• pdftotext (from poppler) — install via brew install poppler or apt install poppler-utils if missing
• Google Fonts (loaded via CDN in the generated HTML)

Phase 1: Paper Deep-Read#

Step 1.0 — Resolve input source#

Determine the input type and obtain a local PDF:

InputActionLocal path ( /path/to/paper.pdf)Use directlyarXiv abs URL ( https://arxiv.org/abs/2603.02096)Convert to https://arxiv.org/pdf/2603.02096 and downloadarXiv PDF URL ( https://arxiv.org/pdf/2603.02096)Download directlyOther PDF URLDownload directly

Download command:

bash

curl -L -o /tmp/paper.pdf "<pdf_url>"

Use the arXiv ID or paper title as the filename when possible (e.g., /tmp/2603.02096_FluxMem.pdf).

Step 1.1 — Extract full text#

bash

pdftotext "<paper.pdf>" /tmp/paper_text.txt

Read the extracted text completely. For papers with critical figures, try extracting key pages as images via pdftoppm and analyzing them with the image tool.

Step 1.2 — Comprehensive analysis#

Create temp_analysis.md extracting:

• Research question, hypotheses, methodology, data sources
• Core findings with key quantitative results
• Theoretical contributions and practical implications
• The paper's fundamental contradiction/gap, novel angle, and method innovation

This step is mandatory — it ensures report quality through structured thinking.

Step 1.3 — Write dual-mode report#

Generate [PaperName]_研读报告.md containing both parts separated by ---:

• Part A — Read part-a-template.md before writing. Deep academic report: structured abstract → introduction → methodology → results → discussion → conclusion → core references.
• Part B — Read part-b-template.md before writing. Core logic extraction: four key elements table → method formula → one-line summary (expert + layperson versions).

Writing standards: Use complete paragraphs over bullet lists. Provide bilingual terms (中英) on first mention. Support every claim with specific data from the paper.

Step 1.4 — Deliver report#

Present the report file and briefly summarize the paper's core innovation, key findings, and theoretical value.

Phase 2: Slides Generation#

If the user only requested a report, stop after Phase 1.

Step 2.0 — Confirm parameters#

Ask the user (all at once):

1. Purpose: 组会汇报 / 学术会议 / 教学 / 其他
2. Length: Short (5-10) / Medium (10-20) / Long (20+)
3. Theme: 亮色 / 暗色 / 偏好的风格
4. Editing: 是否需要浏览器内编辑功能

If the user already specified these in the initial request, skip asking and proceed.

Step 2.1 — Choose style#

Select an appropriate style based on user preferences. Read slide-styles.md for available presets and CSS specifications. Academic presentations generally suit Swiss Modern (light), Notebook Tabs (light), or Bold Signal (dark).

Step 2.2 — Plan slide outline#

Map the report content to a slide structure. Typical academic paper outline (15-19 slides):

SectionSlidesContentTitle1Paper title, authors, affiliations, dateProblem1-2Background, existing gaps, key insightMethod (divider)1Section divider slideMethod details3-5Core modules, formulas, architectureResults (divider)1Section divider slideQuantitative results2-3Main results tables, metrics, comparisonsEfficiency / Ablation1-2Efficiency gains, ablation studiesDiscussion (divider)1Section divider slideAdvantages & Limitations1-2Key properties, future directionsSummary1One-line takeawayThanks1Links, navigation hints

Step 2.3 — Generate HTML#

Read slide-template.md for the mandatory HTML architecture, viewport CSS, JS controller, and inline editing implementation.

Non-negotiable rules:

• Single self-contained HTML file, all CSS/JS inline
• Every .slide must have height: 100vh; height: 100dvh; overflow: hidden
• ALL font sizes and spacing must use clamp() — never fixed px/rem
• Load fonts from Google Fonts or Fontshare — never system fonts
• Include prefers-reduced-motion support
• Max content per slide: 1 heading + 4-6 bullets OR 1 heading + 2 paragraphs. Overflow → split slides.

If inline editing is enabled, implement via contentEditable API + localStorage persistence. Use JS-based hover with 400ms delay for the edit button (not CSS ~ sibling selector — it breaks due to pointer-events). See the template for the complete implementation.

Step 2.4 — Open and deliver#

bash

open [filename].html

Tell the user: file location, slide count, navigation (arrow keys / space / dots), and editing instructions (press E) if enabled.

Both Phases Together#

When the user requests both report and slides, run Phase 1 fully first, then Phase 2. The report serves as the structured source for slide content — do not re-analyze the paper.