Paper-to-Beamer

Convert an academic ML/AI paper (LaTeX source in paper/) into a compilable Beamer presentation, backed by the claude-code-my-workflow infrastructure (agents, quality gates, review pipeline).

Quick start

User: make slides from my paper in paper/
→ Phase 0: scaffold workflow if missing
→ Phase 1-5: read → generate → compile → score → review

Core workflow

Phase 0: Scaffold Workflow Infrastructure (unless --skip-setup)

Before generating slides, verify the project has the claude-code-my-workflow infrastructure. Check for sentinel files:

.claude/agents/slide-auditor.md
.claude/rules/quality-gates.md
scripts/quality_score.py
Preambles/header.tex

If any are missing, scaffold from the bundled scaffold/ directory:

SKILL_DIR=".claude/skills/paper-to-beamer"
# Create target directories
mkdir -p .claude/agents .claude/rules scripts Preambles quality_reports/plans Slides paper
# Copy agents
cp "$SKILL_DIR/scaffold/agents/slide-auditor.md" .claude/agents/
cp "$SKILL_DIR/scaffold/agents/domain-reviewer.md" .claude/agents/
cp "$SKILL_DIR/scaffold/agents/proofreader.md" .claude/agents/
# Copy rules
cp "$SKILL_DIR/scaffold/rules/quality-gates.md" .claude/rules/
cp "$SKILL_DIR/scaffold/rules/no-pause-beamer.md" .claude/rules/
cp "$SKILL_DIR/scaffold/rules/content-invariants.md" .claude/rules/
# Copy scripts and Preambles
cp "$SKILL_DIR/scaffold/scripts/quality_score.py" scripts/
cp "$SKILL_DIR/scaffold/Preambles/header.tex" Preambles/
# Create CLAUDE.md if missing
test -f CLAUDE.md || cat > CLAUDE.md << 'CLAUDE_EOF'
# 论文转Beamer幻灯片专用配置
## 目标
核心任务：读取 `paper/` 目录下的 LaTeX 论文源码，生成一个完整的、可直接编译的学术 Beamer 演示文稿。
## 核心指令
- 读取与分析：深度理解论文的贡献、方法、实验与结论。
- 幻灯片结构：按"标题页→动机→方法→实验→结论→致谢"的顺序规划内容。
- 内容原则：每页不超过 4 个要点，每个要点不超过 6 个单词；多用图表公式，少用文字段落。
- 视觉风格：使用 `metropolis` Beamer 主题，深蓝色调。
- 编译与验证：生成可编译的 `beamer.tex`，并调用 `latexmk -xelatex` 进行编译；如有错误，自动修复。
CLAUDE_EOF

Report: "Workflow scaffolded from skill bundle: 3 agents, 3 rules, quality_score.py, Preambles/header.tex" (plus CLAUDE.md if newly created).

The bundled files are exact copies from claude-code-my-workflow — no LLM generation, no drift. The domain-reviewer is pre-customized for ML/AI papers.

Phase 1: Read & Understand

1. Read the paper main file and all \input{}-ed sections.
2. Extract: contributions, method name & intuition, key equations, main result table/figures, ablation highlights, conclusions.
3. Note: paper title, authors, affiliations, venue, year.

Phase 2: Generate Beamer

Follow the slide structure: Background (optional) → Method → Experiments → Conclusion.

Write Slides/beamer.tex with this preamble baseline:

\documentclass[10pt,aspectratio=169]{beamer}
\usetheme[progressbar=frametitle,block=fill]{metropolis}
\usefonttheme{professionalfonts}  % no Fira Sans dependency
% Colors: deep blue #012169, gold #B9975B, light bg #E8EDF5
% frame title bg=deepblue, fg=white; structure=deepblue

Content conventions (see REFERENCE.md for full details):

• Max 4 bullet points per slide, each ≤6 words
• Visible author-year citations: always (Author et al., Year) — never \cite{key}
• Symbol legends: every new symbol explained on first use in {\tiny ...} below
• Concrete examples for abstract claims (e.g., "Therefore" vs "6" for style vs math tokens)
• Method slides: each baseline method cited with author-year + the year it was proposed
• No \pause, \onslide, or overlay commands
• Only Chinese if user explicitly requests; default English

Phase 3: Compile & Fix

cd Slides && xelatex -interaction=nonstopmode beamer.tex
xelatex -interaction=nonstopmode beamer.tex  # 2-pass, no bibtex needed

Check beamer.log for Error and undefined. Fix any issues, recompile.

Phase 4: Quality Gate

python scripts/quality_score.py Slides/beamer.tex

• Score < 80: fix blocking issues, recompile, re-score
• Score ≥ 80: ready

Phase 5: Review (unless --skip-review)

Run review agents:

1. slide-auditor → overflow, font, spacing
2. proofreader → grammar, typos
3. domain-reviewer → ML claims, equation correctness

Fix critical issues, recompile, re-confirm score ≥ 80.

Adding background (--add-background)

When the audience may not know post-training methods, insert background slides between title and method covering:

1. Pipeline overview — TikZ diagram: Pre-training → SFT/RL/Distill → Inference
2. Evolution story — chain diagram: SFT → RLHF/DPO → GRPO → On-Policy Distill → OPSD, each with the problem it fixed
3. Each method — core idea, formula, I/O, pros/cons block, concrete example, author-year citation
4. Bridge slide — "Why Self-Distillation?" leading into the paper's method

Each background method slide must include: TikZ flow diagram + formula + Pros/Cons blocks + citation.

Triggers

Invoke this skill when user says any of:

• "make slides from this paper"
• "generate beamer / presentation"
• "paper to slides / ppt"
• "create slides for my talk / group meeting"
• "I need to present this paper"

See also

• REFERENCE.md — full symbol checklist, color palette, content invariants
• /slide-excellence — multi-agent comprehensive slide review (post-generation)
• /compile-latex — XeLaTeX 3-pass compilation (if bibtex needed)
• scripts/quality_score.py — quality scoring (80/90/95 thresholds)