Paper Presenter Skill

Purpose

Bridge between paper-parser (MinerU output) and beamer-skill. Analyzes parsed paper content + optional source code to produce a structured "materials package" that feeds directly into beamer-skill's create [topic] workflow.

When to Use

• After running paper-parser to get structured markdown/figures
• User asks to "make a presentation", "create slides", or "summarize this paper as a talk"
• User asks to "present the reproduction results" or "make a comparison between paper and code"

How to Use

Step 1: Ensure Prerequisites

• Parsed output from paper-parser (directory with full.md, images/, JSON files)
• (Optional) Source code from paper-downloader

Step 2: Run the Preparation Script

Standard paper overview:

python skills/paper-presenter/prepare_presentation.py workspace/<paper>/parsed/ -o workspace/<paper>/presentation_materials.md

With source code analysis (paper ↔ code comparison):

python skills/paper-presenter/prepare_presentation.py workspace/<paper>/parsed/ \
    --code-dir workspace/<paper>/code/<repo>/ \
    --style reproduction-report \
    -o workspace/<paper>/presentation_materials.md

Options:

• --style overview — Standard paper overview (~15-20 slides)
• --style deep-dive — Technical deep-dive (~30-40 slides)
• --style reproduction-report — Paper + code + reproduction results comparison
• --code-dir <dir> — Source code repo for paper-code alignment analysis
• --paper-info <file> — paper_info.json for metadata (auto-detected if in standard layout)
• --language en|zh — Output language

Step 3: Feed Materials to beamer-skill

The output presentation_materials.md is designed as Phase 0 input for beamer-skill's create workflow.

Tell the agent:

Create a beamer presentation based on the materials in workspace/<paper>/presentation_materials.md

The beamer-skill will:

1. Read the materials file (Phase 0)
2. Ask the user about duration, audience, scope (Phase 1)
3. Build a slide structure plan (Phase 2)
4. Draft slides using extracted figures and formulas (Phase 3-4)
5. Run quality checks (Phase 5)

Key Features for Reproduction Presentations

When --style reproduction-report and --code-dir are used:

• Code-Theory Alignment Table — maps paper sections to source code files
• Comparison Slide Suggestions — paper formula on left, code on right
• Implementation Detail Highlights — what's in code but not in paper
• Delta Analysis Structure — original results vs reproduced results

Dependencies

• Python 3.10+
• No external libraries required (uses only stdlib + json)
• Requires: beamer-skill for the actual presentation generation
  • Bundled at: ../beamer-skill/beamer/ in this monorepo
  • Original: https://github.com/Noi1r/beamer-skill (MIT License by Noi1r)
  • To install: copy or symlink beamer-skill/beamer/ to your project's .gemini/skills/beamer/

Output Format

The output is a structured Markdown file containing:

1. Paper metadata (title, authors, venue)
2. Section-by-section content analysis
3. Available figures with paths
4. Key mathematical formulas
5. Source code analysis (if provided)
6. Code-theory alignment mapping (if code provided)
7. Suggested slide structure for the chosen style
8. Instructions for beamer-skill agent