aipoch/scientific-schematics

Scientific Schematics Skill

When to Use

• Creating journal-ready figures (clean typography, consistent styling, high resolution) from a short textual description.
• Producing poster-friendly diagrams that prioritize readability at distance (larger labels, stronger contrast).
• Drafting neural network architecture schematics (e.g., Transformer blocks, attention modules) for papers or slides.
• Generating biological pathway visuals (e.g., Krebs cycle) with iterative quality review.
• Rapidly iterating on a diagram concept when you need AI-assisted refinement loops instead of manual redraws.

Key Features

• Text-to-diagram automation: Converts a natural-language prompt into a publication-quality schematic.
• Iterative generate → review → refine loop: Automatically improves the figure until a quality threshold is met.
• Document-type aware critique: Reviewer feedback adapts to journal vs poster requirements.
• Model-configurable pipeline: Choose separate LLMs for generation and vision-based review.
• Output validation: Performs final checks (e.g., resolution/accessibility considerations) before saving to figures/.
• Reference guidance:
  • Best practices: references/best_practices.md
  • Supported diagram categories: references/diagram_types.md

Dependencies

• Python 3.10+ (recommended)
• Python packages:
  • pillow (PIL)
  • matplotlib
  • requests
• Environment:
  • OPENROUTER_API_KEY (required)

Example Usage

1) Set the OpenRouter API key

Windows (PowerShell)

$env:OPENROUTER_API_KEY="your_key_here"

Linux/macOS

export OPENROUTER_API_KEY="your_key_here"

2) Run the generator (journal/poster)

python scripts/generate_schematic.py "Transformer architecture with attention mechanism" --doc-type journal

3) Override the generation model

python scripts/generate_schematic.py "Krebs cycle" --doc-type journal --generator anthropic/claude-3.5-sonnet

4) (Optional) Override both generator and reviewer

python scripts/generate_schematic.py "Flowchart of a clinical trial enrollment pipeline" \
  --doc-type poster \
  --generator google/gemini-2.0-flash-001 \
  --reviewer google/gemini-2.0-flash-001

Implementation Details

Pipeline Stages

1. Generation

   • A code-capable LLM converts the prompt into a diagram image.
   • Default generator model: google/gemini-2.0-flash-001.

2. Review

   • A vision-capable LLM evaluates the generated image against the target --doc-type.
   • Default reviewer model: google/gemini-2.0-flash-001.
   • The reviewer returns actionable critique and a numeric quality score.

3. Refinement Loop

   • If the score is below the acceptance threshold (e.g., 8.5/10), the system re-enters generation using the reviewer’s feedback as constraints.
   • This repeats until the threshold is met or the run terminates by internal stopping conditions.

4. Finalization

   • Performs final checks such as resolution suitability and accessibility-oriented considerations (e.g., legibility).
   • Saves the final artifact to the figures/ directory.

Key Parameters

• --doc-type <journal|poster>: Controls review criteria (e.g., density/precision for journals vs readability/scale for posters).
• --generator <model_id>: Model used to produce the diagram.
• --reviewer <model_id>: Model used to critique the diagram.
• Quality threshold: A numeric cutoff (example: 8.5/10) that determines whether refinement continues.