PaperBanana Skills for Claude Code

One sentence in, publication-quality academic figure out.
Powered by a 5-agent pipeline that plans, styles, generates, and self-critiques your illustrations.

English | 中文

Gallery

Biology — Signal Pathway	NLP — RAG Pipeline
Data Engineering — Lakehouse	Medical AI — U-Net + Mamba
Medical Imaging — TextMamba3D gpt-image-2 · paper-grade info density	Game Theory — Influence Diagram Gemini · soft pastel academic aesthetic
Ablation Study — BraTS 2023 gpt-image-2 · 2×2 MRI panels + Dice bar chart	Scientific Slide — scRNA-seq Workflow paperbanana-slide-deck · single-cell analysis pipeline

All figures generated from plain text descriptions — zero manual drawing.

Slide Deck Showcase — “The Flywheel Learning Method”

A real 10-slide lecture deck built with paperbanana-slide-deck. Below: 4 selected slides showing set-wide style consistency (same warm off-white palette, sketch-notes hand-drawn typography, and gear motif across the whole deck).

Slide 1 — Cover	Slide 4 — Flywheel Model
Slide 7 — AI Tools Do's and Don'ts	Slide 10 — Let the Flywheel Spin

One command: paperbanana-slide-deck picks a style preset, plans the outline, drafts per-slide prompts, then generates all slides with consistent design tokens.

More Examples (architecture diagrams, traditional aesthetics)

Transformer Architecture	Mamba SSM Architecture
RAG Pipeline
Chinese Calligraphy — 自律 (Self-Discipline) Gemini · bold expressive brushwork + 飞白 on xuan paper

Skills in this Marketplace

Feature Matrix

What’s New in v4.3 — GPT Image 2 First-Class Support

OpenAI released gpt-image-2 on 2026-04-21. PaperBanana v4.3 integrates it natively so the full Retriever → Planner → Stylist → Visualizer → Critic pipeline runs on gpt-image-2 outputs. You get quality-gated images at up to 2048×1152 without leaving paperbanana.

Adapter upgrade

Switching is a two-flag change:

python -m paperbanana.cli generate \
  --image-provider openai --image-model gpt-image-2 \
  --aspect-ratio 16:9 \
  --input prompt.txt --caption "..."

Auto routing by scenario

The skill picks the right provider based on your request’s signal:

Routing is calibrated from a 16-prompt controlled comparison (details: docs/superpowers/specs/2026-04-23-image-router-design.md in the companion repo).

Before / After — routing in action

These pairs come from the same prompt sent to both providers. The routing table exists because each model has specific strengths and specific bugs.

1. Chinese slide titles — GPT wins (Gemini has a duplicate-character bug)

Gemini Title reads "飞轮模飞轮模型" — the prefix "飞轮模" is duplicated. Not viable for slide decks.

gpt-image-2 Title renders cleanly: "飞轮模型 — 核心概念". Routing sends Chinese slides here.

2. Semantic correctness (diffusion process) — GPT wins

Gemini Cat images at x_0 through x_4 look identical; only x_T is noise. Semantics and visuals don't match.

gpt-image-2 Cat actually degrades step-by-step — visually faithful to the diffusion process.

3. Traditional Chinese calligraphy — Gemini wins (bolder brushwork)

Gemini Bold expressive strokes with visible 飞白 and xuan-paper fiber — the prompt asked for "bold" and got it.

gpt-image-2 Technically correct characters but the stroke feels restrained. The routing sends 书法/山水/古风 prompts to Gemini.

Verdict

You don’t need to know any of this — just ask for a figure and paperbanana picks. Or override with --image-provider openai|gemini|both. The Critic loop runs on whatever the pipeline picks, so quality stays gated regardless.

What’s New in v4.0

Eval-First Quality System

A 6-item binary checklist evaluator that measures academic figure quality without human reference images:

Baseline: 76% → 100% after prompt optimization. Color restraint was the bottleneck (33% → 100%).

Autoresearch Self-Optimization

Automated prompt mutation loop inspired by Karpathy’s autoresearch:

Mutate prompt → Generate figures → Evaluate checklist → Keep or Revert → Repeat

• One mutation per round (isolation principle)
• Targets weakest checklist dimension automatically
• Versioned prompt snapshots + JSONL changelog
• Stop condition: 3 consecutive rounds at 90%+ or 20 rounds max

Multi-Venue Academic Styles

/paperbanana generate method.txt "Architecture overview" --venue neurips

Built-in style guides for NeurIPS, ICML, ACL, IEEE — each with venue-specific color palettes, layout conventions, and typography.

Robust Error Handling

Quick Start

# 1. Install PaperBanana
git clone https://github.com/llmsresearch/paperbanana.git
cd paperbanana && pip install -e ".[google]"

# 2. Add the marketplace & install skills
claude plugin marketplace add PlutoLei/paperbanana-skill
claude plugin install paperbanana@paperbanana-skills
claude plugin install paperbanana-slide-deck@paperbanana-skills --scope project  # optional

# 3. Generate your first figure
# /paperbanana A 4-layer CNN with batch normalization for image classification

Note: This repository contains Claude Code skill definitions (SKILL.md files). The underlying Python package lives at llmsresearch/paperbanana.

Why PaperBanana?

Pipeline Architecture

The pipeline runs iteratively: the Critic evaluates each output against academic quality criteria and either accepts it or sends revision instructions back to the Planner. Parse failures are handled safely — never silently approved.

Slide Deck Orchestrator

End-to-end presentation creation: analyze content → select from 23 visual styles → generate outlines → batch-generate 4K slides → merge to PPTX/PDF.

Commands

# Generate with venue-specific style
/paperbanana generate method.txt "Overview of the proposed framework" --venue neurips --optimize

# Generate from PDF
/paperbanana generate paper.pdf "Architecture diagram" --pages 3-5

# Auto-refine until Critic is satisfied
/paperbanana generate method.txt "Pipeline overview" --auto

# Continue with feedback
/paperbanana generate --continue --feedback "Make the arrows thicker and add color coding"

# Custom provider and aspect ratio
/paperbanana generate method.txt "Wide pipeline" --vlm-provider anthropic --aspect-ratio 16:9

# Batch generate slides with style
/paperbanana slide-batch prompts/ --resolution 4k --style ml-ai --iterations 3

Supported Providers

Retry policy: Transient errors (429, 5xx) retry with exponential backoff. Auth errors (401, 403) fail immediately — no wasted retries.

Installation

Option A: Plugin marketplace (recommended)

claude plugin marketplace add PlutoLei/paperbanana-skill
claude plugin install paperbanana@paperbanana-skills
claude plugin install paperbanana-slide-deck@paperbanana-skills --scope project  # optional

Option B: Manual install

# paperbanana skill (user-level)
mkdir -p ~/.claude/skills/paperbanana
curl -o ~/.claude/skills/paperbanana/SKILL.md \
  https://raw.githubusercontent.com/PlutoLei/paperbanana-skill/master/plugins/paperbanana/skills/paperbanana/SKILL.md

# paperbanana-slide-deck skill (project-level, optional)
mkdir -p .claude/skills/paperbanana-slide-deck
curl -o .claude/skills/paperbanana-slide-deck/SKILL.md \
  https://raw.githubusercontent.com/PlutoLei/paperbanana-skill/master/plugins/paperbanana-slide-deck/skills/paperbanana-slide-deck/SKILL.md

PaperBanana package setup

git clone https://github.com/llmsresearch/paperbanana.git
cd paperbanana
pip install -e ".[google]"          # Gemini (default, free tier available)
# pip install -e ".[all]"           # All providers
python -m paperbanana.cli setup     # Interactive API key configuration

Style Presets (23 available)

Use --style <name> with slide or slide-batch.

Evaluation Infrastructure

PaperBanana v4.0 includes a complete evaluation system for measuring and improving figure quality:

evaluation/
├── checklist.py          # 6-item binary pass/fail evaluator
├── judge.py              # VLM-as-Judge comparative evaluation
├── benchmark.py          # End-to-end benchmark harness
└── prompt_ablation.py    # A/B prompt comparison runner

scripts/
├── run_checklist_baseline.py   # Run checklist on existing outputs
└── autoresearch_loop.py        # Automated prompt optimization

Run your own baseline:

python scripts/run_checklist_baseline.py --output-dir outputs/ --report baseline.json

Run autoresearch optimization:

python scripts/autoresearch_loop.py --test-inputs data/checklist_test_set --max-rounds 10 --target 90

Troubleshooting

Contributing

Contributions welcome! See the Contributing Guide.

License

MIT