Engineering Figure Agent

Preview

Why It Exists

Most figure tools treat every paper figure as the same image prompt problem. Engineering papers do not work that way.

Engineering Figure Agent is intentionally lighter than a full paper-upload platform. It is built for researchers who already know what claim a figure should support and want a cleaner production path inside an agent workflow.

Quick Start

Install and run the setup check:

& "$HOME/.codex/skills/engineering-figure-agent/scripts/install_and_test.ps1" -RunSetupCheck

Open the wizard:

& "$HOME/.codex/skills/engineering-figure-agent/scripts/wizard.ps1"

Or generate a conceptual figure directly:

python "$HOME/.codex/skills/engineering-figure-agent/scripts/generate_image.py" `
  --figure-template system-architecture `
  --lang en `
  "A retrieval-augmented generation system with OCR, chunking, embedding, vector search, reranking, and answer synthesis."

For installation details, use the language-specific guides:

• README.zh-CN.md: Windows 最短安装路径、验证方法、中文示例
• README.en.md: setup, verification, examples, positioning

Two Modes

Image Mode

Use image mode when visual structure matters more than exact numeric geometry.

Best for:

• system architecture diagrams
• algorithm workflows
• graphical abstracts
• electronics or embedded-system schematics
• reference-inspired redraws and layout exploration

Supported conceptual-image backends:

Example:

python "$HOME/.codex/skills/engineering-figure-agent/scripts/generate_image.py" `
  --provider openai `
  --model gpt-image-1.5 `
  --figure-template system-architecture `
  --lang en `
  --openai-quality auto `
  --openai-size auto `
  "A retrieval-augmented generation system with OCR, chunking, embedding, vector search, reranking, and answer synthesis."

Plot Mode

Use plot mode when values, axes, and geometry must stay exact.

Best for:

• benchmark bar charts
• ablation plots
• trend curves
• heatmaps
• scatter plots
• multi-panel quantitative figures

Rule of thumb:

Workflow

Recommended upstream handoff:

1. Use ai-research-writing-guide to decide what claim the figure should support.
2. Write or collect the figure brief: audience, claim, modules, labels, data, and target style.
3. Use engineering-figure-agent to render the diagram or exact plot.
4. Verify labels, numeric truth, publication style, and export format before paper submission.

Platform Adapters

Core platform-neutral contracts:

• Figure brief spec
• Plot request schema
• Figure brief schema

Support Matrix

Repository Map

What It Is Not

• Not a full paper-upload web platform.
• Not a replacement for checking scientific truth, labels, and numeric values.
• Not a single prompt that treats plots, diagrams, and schematics as the same task.
• Not a place to commit real API keys or private provider relay details.

Notes

• Keep real API keys outside the repository.
• Prefer local plotting for exact quantitative figures.
• Keep provider-specific private relay details out of public docs unless clearly marked as optional examples.
• For exact publication plots, never rely on image generation for the values, axes, or benchmark geometry.

License

MIT License. See LICENSE.