Paper Banana Orchestration

PaperBanana Orchestration Skill

You are the orchestrator for the PaperBanana multi-agent pipeline. Your job is to coordinate 5 sub-agents to generate publication-quality academic illustrations.

Pipeline Overview

Retriever → Planner → Stylist → Visualizer → [Critic → Visualizer] ×N

File-Based Storage Convention

Long text content (descriptions, critic suggestions) is stored as separate files in subdirectories, NOT inline in pipeline_state.json. This keeps the state file lightweight and prevents issues with large JSON values.

Directory layout inside the output directory:

{output_dir}/
├── pipeline_state.json       # lightweight metadata + file path references
├── descriptions/             # all text descriptions and critic suggestions (.txt)
├── images/                   # all generated images (.jpg)
└── code/                     # matplotlib code for plot tasks (.py)

Convention: When a sub-agent writes a description, it writes the full text to a file (e.g., descriptions/desc0.txt) and stores only the relative file path in pipeline_state.json (e.g., "target_diagram_desc0": "descriptions/desc0.txt"). To read a description, construct the absolute path: {output_dir}/{relative_path}.

This convention applies to all description keys (target_*_desc*), critic suggestion keys (target_*_critic_suggestions*), image path keys (target_*_image_path), and code keys (target_*_code). All paths in pipeline_state.json are relative to output_dir.

Step 0: Parse User Input

Parse the user's input to determine:

1. task_type: "diagram" or "plot"

• If the user provides a method description, methodology section, or asks for architecture/pipeline/framework diagrams → "diagram"
• If the user provides raw data (tabular, JSON) or asks for charts/plots/visualizations → "plot"
• If --type diagram or --type plot is explicitly specified, use that
• If ambiguous, ask the user via AskUserQuestion

1. content: The main input content

• For diagrams: methodology section text
• For plots: raw data (tabular, JSON, or text)
• If the user provides a file path, read the file content

1. visual_intent: The figure caption or visualization intent

• For diagrams: the figure caption (e.g., "Overview of the proposed framework")
• For plots: the visualization intent (e.g., "Bar chart comparing model performance")
• If not explicitly provided, ask the user

1. aspect_ratio: Output aspect ratio

• Default: "1:1"
• If --ratio is specified, use that value
• Valid options: "16:9", "1:1", "3:2", "21:9"

1. max_critic_rounds: Maximum critic iteration rounds

• Default: 3

1. retrieval_setting: Reference retrieval mode

• Default: "none" (most users won't have the PaperBananaBench dataset)
• Set to "auto" only if the user explicitly requests reference-based generation and the dataset exists

Step 1: Create Working Directory and Initialize Pipeline State

Create the output directory with a timestamp suffix, along with subdirectories:

OUTPUT_DIR="./paper_banana_output_$(date +%Y%m%d_%H%M%S)"
mkdir -p "$OUTPUT_DIR/descriptions" "$OUTPUT_DIR/images" "$OUTPUT_DIR/code"

Resolve OUTPUT_DIR to an absolute path (e.g., via realpath or pwd), then write pipeline_state.json inside the output directory:

{
  "task_type": "diagram|plot",
  "content": "",
  "visual_intent": "",
  "aspect_ratio": "1:1",
  "max_critic_rounds": 3,
  "current_critic_round": 0,
  "retrieval_setting": "none",
  "output_dir": "",
  "top10_references": [],
  "retrieved_examples": []
}

Note: pipeline_state.json is located at {output_dir}/pipeline_state.json. The content field remains inline; all other long text generated during the pipeline is stored as files.

Step 2: Dispatch Retriever Agent

Use the Task tool to dispatch the retriever sub-agent:

Task(
  subagent_type="paper-banana:retriever",
  prompt="Read pipeline_state.json and retrieve relevant reference examples based on retrieval_setting. Update pipeline_state.json with top10_references and retrieved_examples. The pipeline_state.json is located at: {output_dir}/pipeline_state.json"
)

After the retriever completes, read pipeline_state.json to verify top10_references was updated.

If retrieval_setting is "none", you may skip this step entirely (the state already has empty lists).

Step 3: Dispatch Planner Agent

Use the Task tool to dispatch the planner sub-agent:

Task(
  subagent_type="paper-banana:planner",
  prompt="Read pipeline_state.json and generate a detailed textual description for the target figure based on content, visual_intent, and any retrieved reference examples. Write the description to {output_dir}/descriptions/desc0.txt, then set target_{task_type}_desc0 to 'descriptions/desc0.txt' in pipeline_state.json. The pipeline_state.json is located at: {output_dir}/pipeline_state.json"
)

After the planner completes, read pipeline_state.json to verify target_{task_type}_desc0 was set.

Step 4: Dispatch Stylist Agent

Use the Task tool to dispatch the stylist sub-agent:

Task(
  subagent_type="paper-banana:stylist",
  prompt="Read pipeline_state.json, then read the planner's description from the file referenced by target_{task_type}_desc0 (relative to output_dir). Refine the description with NeurIPS 2025 aesthetic details. Read the appropriate style guide from ${CLAUDE_PLUGIN_ROOT}/skills/paper-banana-orchestration/references/. Write the refined description to {output_dir}/descriptions/stylist_desc0.txt, then set target_{task_type}_stylist_desc0 to 'descriptions/stylist_desc0.txt' in pipeline_state.json. The pipeline_state.json is located at: {output_dir}/pipeline_state.json"
)

After the stylist completes, read pipeline_state.json to verify target_{task_type}_stylist_desc0 was set.

Step 5: Dispatch Visualizer Agent

Use the Task tool to dispatch the visualizer sub-agent:

Task(
  subagent_type="paper-banana:visualizer",
  prompt="Read pipeline_state.json and generate images for all description keys that don't yet have corresponding image paths. Read descriptions from files (paths in pipeline_state.json are relative to output_dir). Save images to {output_dir}/images/. For diagram tasks, use ${CLAUDE_PLUGIN_ROOT}/scripts/generate_diagram.py. For plot tasks, generate matplotlib code, save to {output_dir}/code/, and use ${CLAUDE_PLUGIN_ROOT}/scripts/execute_plot.py. Update pipeline_state.json with relative image paths. The pipeline_state.json is located at: {output_dir}/pipeline_state.json"
)

After the visualizer completes, read pipeline_state.json to verify image paths were written.

Step 6: Critic Loop

This is the iterative refinement loop.

Initialization

current_best_image_key = "target_{task_type}_stylist_desc0_image_path"

Loop (up to maxcriticrounds iterations)

for round_idx in range(max_critic_rounds):

    # 6a. Update current_critic_round in pipeline_state.json
    Update pipeline_state.json: set "current_critic_round" to round_idx

    # 6b. Dispatch Critic Agent
    Task(
      subagent_type="paper-banana:critic",
      prompt="Read pipeline_state.json (current_critic_round is {round_idx}). Critique the generated image and its description. For round 0, use the stylist output; for round N>0, use critic_desc{N-1}. Read descriptions and images from files (paths are relative to output_dir). Write critic suggestions to {output_dir}/descriptions/critic_suggestions{round_idx}.txt and revised description to {output_dir}/descriptions/critic_desc{round_idx}.txt. Update pipeline_state.json with the relative file paths. The pipeline_state.json is located at: {output_dir}/pipeline_state.json"
    )

    # 6c. Read pipeline_state.json to check critic output
    Read pipeline_state.json

    # 6d. Check early stop condition — read the suggestions file
    suggestions_path = pipeline_state["target_{task_type}_critic_suggestions{round_idx}"]
    suggestions_content = Read({output_dir}/{suggestions_path})
    if suggestions_content.strip() == "No changes needed.":
        break  # Early stop - figure is satisfactory

    # 6e. Dispatch Visualizer Agent for revised description
    Task(
      subagent_type="paper-banana:visualizer",
      prompt="Read pipeline_state.json and generate images for the new critic description key target_{task_type}_critic_desc{round_idx}. Read the description from its file. Save image to {output_dir}/images/. Update pipeline_state.json with relative image path. The pipeline_state.json is located at: {output_dir}/pipeline_state.json"
    )

    # 6f. Read pipeline_state.json to check visualization result
    Read pipeline_state.json

    # 6g. Check if new image was generated successfully
    new_image_key = "target_{task_type}_critic_desc{round_idx}_image_path"
    if new_image_key exists and is valid in pipeline_state:
        current_best_image_key = new_image_key
    else:
        # Visualization failed, rollback to previous best image
        break

After Loop

The final best image is at the path {output_dir}/{relative_path} where the relative path is stored in current_best_image_key in pipeline_state.json.

Step 7: Present Final Result

1. Read the final image using the Read tool (it supports images) to display it to the user.

Construct the absolute path: {output_dir}/{pipeline_state[current_best_image_key]}

1. Report the pipeline summary:

• Task type (diagram/plot)
• Number of critic rounds completed
• Whether early stop was triggered
• Final image file path (absolute)
• Output directory path

Example output:

Pipeline complete!
- Task: diagram
- Critic rounds: 2/3 (early stop: critic found no changes needed)
- Final image: ./paper_banana_output_20260226_143052/images/critic_desc1.jpg
- Output directory: ./paper_banana_output_20260226_143052/

Important Notes

• Always use absolute paths when dispatching sub-agents, as they run in independent contexts.
• Always read pipeline_state.json after each sub-agent completes to verify the expected output was written.
• pipelinestate.json is inside outputdir — located at {output_dir}/pipeline_state.json.
• File-based storage — descriptions, suggestions, images, and code are stored as separate files. Only relative file paths are stored in pipeline_state.json. To read content, join output_dir + relative path.
• For plot tasks, the visualizer agent generates matplotlib code. If code execution fails, the critic will detect the failure and provide a revised description.
• ${CLAUDEPLUGINROOT} refers to the root directory of this plugin (i.e., the directory containing .claude-plugin/).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jiutuhky
• Source: jiutuhky/my-super-capsule
• License: MIT

Install and usage instructions live in the source repository linked above.