Agent-native figure production for engineering and CS papers.
把工程论文里的系统架构图、算法流程图、实验曲线和多面板图，拆成可控的生成流程：概念图走 image mode，精确数值图走 plot mode。
| System / Architecture | Cooperative Perception | Safety Taxonomy |
|---|---|---|
 |
 |
 |
| Dense Systems Diagram | Deployment Scenario |
|---|---|
 |
 |
| Exact Local Plot |
|---|
 |
Most figure tools treat every paper figure as the same image prompt problem. Engineering papers do not work that way.
| Figure need | Better path |
|---|---|
| System architecture, pipeline, schematic, graphical abstract | image mode with an engineering-aware prompt template |
| Benchmark curves, ablation bars, heatmaps, scatter plots | plot mode with local plotting and exact values |
| Mixed conceptual + quantitative figure | Render numeric panels locally, then compose or describe the conceptual panels |
| Reference-inspired redraw | Use image mode for structure and style exploration, then manually verify labels and layout |
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.
Install and run the setup check:
& "$HOME/.codex/skills/engineering-figure-agent/scripts/install_and_test.ps1" -RunSetupCheckOpen the wizard:
& "$HOME/.codex/skills/engineering-figure-agent/scripts/wizard.ps1"Or generate a conceptual figure directly:
python "$HOME/.codex/skills/engineering-figure-agent/scripts/generate_image.py" `
--figure-template system-architecture `
--lang en `
"A retrieval-augmented generation system with OCR, chunking, embedding, vector search, reranking, and answer synthesis."For installation details, use the language-specific guides:
- README.zh-CN.md: Windows 最短安装路径、验证方法、中文示例
- README.en.md: setup, verification, examples, positioning
Use image mode when visual structure matters more than exact numeric geometry.
Best for:
- system architecture diagrams
- algorithm workflows
- graphical abstracts
- electronics or embedded-system schematics
- reference-inspired redraws and layout exploration
Supported conceptual-image backends:
| Provider | Use case | Config |
|---|---|---|
gemini / banana |
Google Gemini / Banana-compatible generation | NANOBANANA_* variables |
openai |
OpenAI Image API backend for conceptual figures and edits | OPENAI_API_KEY or OPENAI_API_KEY_FILE |
Example:
python "$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."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:
| If the figure needs... | Use |
|---|---|
| exact values, axes, error bars, or benchmark geometry | plot mode |
| concept explanation, layout exploration, or architecture visuals | image mode |
| both | local plot panels first, image generation only for conceptual parts |
Recommended upstream handoff:
- Use
ai-research-writing-guideto decide what claim the figure should support. - Write or collect the figure brief: audience, claim, modules, labels, data, and target style.
- Use
engineering-figure-agentto render the diagram or exact plot. - Verify labels, numeric truth, publication style, and export format before paper submission.
| Platform | Entry | Best use |
|---|---|---|
| Codex | SKILL.md | Full agent-native workflow with local scripts |
| Claude Code | adapters/claude-code/ | Local repository figure work and prompt/plot generation |
| ChatGPT / Claude web | docs/prompt-pack.md | Chat-only figure brief and prompt drafting |
| VS Code / Obsidian | templates/figure-brief/ | Edit and archive figure briefs, prompts, and plot requests |
Core platform-neutral contracts:
| Platform | Status | Notes |
|---|---|---|
| Windows | tested | primary tested platform; helper scripts supported first |
| macOS | reported working | successful installs have been reported, including AI-assisted setup |
| Linux | expected for core Python workflow | some environments may need small manual adjustments |
| Path | Purpose |
|---|---|
| SKILL.md | Internal Codex skill instructions |
| providers.md | Provider-neutral API configuration notes |
| references/ | Figure templates, plotting rules, prompt templates, API notes |
| scripts/ | Image generation, plotting, setup checks, wizard scripts |
| examples/figure-briefs/ | Reusable figure brief examples |
| docs/examples/ | Showcase images, prompts, and notes |
| templates/figure-brief/ | Platform-neutral figure brief templates |
| adapters/ | Claude Code and future platform adapters |
- 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.
- 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.
MIT License. See LICENSE.