AgentSkillsCN

paperbanana

利用 PaperBanana 多智能体框架，根据文本描述生成具备出版质量的学术图表与统计图。适用于用户希望为学术论文创建方法论图、架构示意图，或生成统计图表时使用。

SKILL.md
--- frontmatter
name: paperbanana
description: Generate publication-quality academic diagrams and statistical plots from text descriptions using the PaperBanana multi-agent framework. Use when the user wants to create methodology diagrams, architecture figures, or statistical plots for academic papers.
argument-hint: [generate|plot|evaluate] [description or file path]
allowed-tools: Read, Write, Bash, Glob, Grep, AskUserQuestion

PaperBanana - Academic Illustration Generator

Generate publication-quality academic diagrams using a multi-agent AI pipeline (Retriever, Planner, Stylist, Visualizer, Critic).

Prerequisites

  1. •PaperBanana must be installed: pip install -e ".[google]" from the paperbanana repo
  2. •Google Gemini API key must be configured in the project's .env file:
    code
    GOOGLE_API_KEY=your-api-key-here
    Get a free key at: https://makersuite.google.com/app/apikey

Environment Detection

Before running commands, detect the PaperBanana installation:

  1. •Try running python -m paperbanana.cli --help to check if paperbanana is available in the current Python environment
  2. •If not found, search for common locations:
    • •Check if a paperbanana directory exists in the current workspace
    • •Look for virtual environments (.venv, venv, conda envs) that may have it installed
  3. •Set the working directory to where PaperBanana is installed (contains paperbanana/ package and .env)

Command pattern:

bash
cd <paperbanana_project_dir> && python -m paperbanana.cli <command> <args>

On Windows with conda, you may need:

code
powershell -Command "Set-Location '<paperbanana_dir>'; & '<python_path>' -m paperbanana.cli <command> <args> 2>&1"

Set timeout to 300000 (5 minutes) for generate/plot commands since they call external APIs.

Usage Modes

Mode 1: generate - Methodology Diagrams

Generate a methodology/architecture diagram from text.

When user provides a file path:

bash
python -m paperbanana.cli generate --input '<file_path>' --caption '<caption>' --iterations 2

When user provides inline text (no file):

  1. •Write the user's methodology text to a temporary file (e.g., temp_input.txt) in the PaperBanana project directory
  2. •Run the generate command with that temp file as --input
  3. •Clean up the temp file after completion

Parameter guidelines:

  • •--caption: Extract or craft a concise figure caption from the user's description. Should describe what the diagram communicates (e.g., "Overview of the proposed encoder-decoder architecture").
  • •--iterations: Default to 2 for balanced quality/speed. Use 1 for quick drafts, 3 for highest quality.
  • •--vlm-model: Default is gemini-3-flash-preview. For complex diagrams, suggest gemini-3-pro-preview.
  • •--output: Optional output path. If not specified, saves to outputs/<run_id>/final_output.png.

Mode 2: plot - Statistical Plots

Generate statistical plots from CSV or JSON data.

bash
python -m paperbanana.cli plot --data '<data_file>' --intent '<intent>' --iterations 2
  • •--data: Path to CSV or JSON file
  • •--intent: What the plot should communicate (e.g., "Bar chart comparing model accuracy across datasets")

Mode 3: evaluate - Diagram Evaluation

Compare a generated diagram against a human reference.

bash
python -m paperbanana.cli evaluate --generated '<gen_path>' --reference '<ref_path>' --context '<text_path>' --caption '<caption>'

Argument Parsing

Parse $ARGUMENTS as follows:

Input PatternAction
generate <file.txt> <caption>Use file as --input, text as --caption
generate <description text>Write text to temp file, auto-generate caption
plot <data.csv> <intent>Use file as --data, text as --intent
evaluate <gen.png> <ref.png>Use as --generated and --reference
Just a description (no subcommand)Default to generate mode

If the user provides only a description without specifying a mode, default to generate.

After Generation

  1. •Parse the output to find the image path (look for "Output saved to:" in stdout)
  2. •Use the Read tool to display the generated image to the user
  3. •Report the Run ID, iteration count, and any Critic feedback
  4. •If the Critic suggested revisions, inform the user they can re-run with more iterations

Examples

User: /paperbanana generate method.txt Overview of sparse attention transformer

Action: Run generate with the provided file and caption.

User: /paperbanana Our model consists of an encoder that processes input features through 3 convolutional layers, followed by a decoder with skip connections

Action: Write text to temp file, generate caption "Overview of the proposed encoder-decoder architecture with skip connections", run generate.

User: /paperbanana plot results.csv Bar chart comparing accuracy of ResNet vs VGG vs EfficientNet

Action: Run plot with the CSV file and the intent description.