Research Drawio Skill

Research Draw.io Skill

中文使用说明

当用户提出“科研流程图”“论文配图”“技术路线图”“机制示意图”“图形摘要”或“draw.io 作图”等需求时触发本 skill。默认产出可编辑的 .drawio 源文件，并按需生成 SVG/PDF/PNG 预览；如果用户只给出主题而未给出完整拓扑，先明确科学信息、图的角色和主要阅读方向，再开始布局。

本 skill 只负责科研图的结构、绘制、审查和导出，不替用户臆造实验结果、统计数值或因果结论。涉及公式、生物结构、模型模块或外部 SVG 时，遵循对应参考文件和导出 QA 规则；任何外部素材的来源与许可应在工作记录中保留。

Build scientific flowcharts as visual arguments, not decorative process maps. The diagram must make a paper's method, mechanism, cohort flow, model architecture, or analytical logic easier to review.

This skill mirrors the nature-figure philosophy: first define the scientific claim and evidence hierarchy, then choose the minimum diagram structure, then engineer the layout and connectors, then apply restrained journal style and export QA.

The output should not be an all-text box diagram when the topic has visual entities. Use editable draw.io primitives to build semantic glyphs for the key scientific objects, while keeping labels short and reviewable.

Routing Protocol

Follow these steps whenever the skill is invoked.

1. Load the manifest and core layer

Read manifest.yaml. Then read every file listed under always_load:

• static/core/contract.md
• static/core/stance.md

Do not draw from memory or from a favorite template before the contract exists. If the diagram contains formulas, load references/math-typesetting.md before authoring any formula labels. If the diagram contains biological structures, charts, tensors, model modules, or other visual entities, load references/composite-elements.md before drawing those entities. If the diagram is traced from a GPT/imagegen/raster reference and close visual matching matters, load references/strict-visual-comparison.md before final QA.

2. Establish the diagram contract before authoring

Create a compact working contract:

1. Scientific message: the one-sentence point the flowchart must defend.
2. Diagram role: method overview, experimental workflow, cohort flow, mechanism, analytical pipeline, graphical abstract, or hybrid figure panel.
3. Topology: ordered nodes, grouped modules, edge meanings, branches, loops, and inputs/outputs.
4. Layout blueprint: rows, columns, modules, gutters, and connector corridors.
5. Composite-element plan: which nodes become grouped visual glyphs, which draw.io primitives compose them, and what each primitive means.
6. Visual vocabulary: colors, shapes, icons, and labels that map to scientific entities rather than decoration.
7. Math contract: formula cells, MathJax delimiters, and fallback plain labels.
8. Export contract: target size, editable source, SVG/PDF needs, and review risks.

If the user provides only a topic, infer a provisional message and topology, then make the uncertainty visible in the working notes or response.

3. Use draw.io as the default editable source

Use diagrams.net/draw.io .drawio as the source format unless the user asks for another format. If no export format is specified, provide or prepare:

• .drawio editable source as the primary artifact.
• .svg as the primary publication/export preview.
• .pdf or .png only when requested or needed for submission/preview.

When editing an existing .drawio, preserve user content, page structure, IDs, grouping, and geometry unless the requested redesign requires a change.

4. Engineer layout before connectors

Before writing edges, create a grid layout plan:

• Place modules on a strict column/row grid with consistent node sizes.
• Decide which major nodes should be visual glyphs instead of plain text boxes.
• Reserve empty routing corridors between modules and around formula blocks.
• Route connectors only through corridors, never through text-bearing shapes.
• Use the simplest valid connector: direct line first, one-bend orthogonal line second, multi-bend detour only when it avoids a real obstacle.
• Prefer separate label cells near connectors over text directly on long edges.
• Avoid crossings by changing geometry before adding style.

For detailed rules, load references/layout-and-routing.md whenever creating or repairing a .drawio file.

5. Build semantic composite elements

Use grouped draw.io primitives when they clarify the science:

• Use bars + axes for a "quantitative result" or "metric comparison" glyph.
• Use paired circles + crossing rails for a DNA/RNA helix glyph.
• Use tiled rectangles for matrices, heatmaps, attention maps, or omics tables.
• Use circles connected by lines for neural networks, graphs, or cell-cell communication.
• Use ellipses/circles with internal marks for cells, nuclei, organelles, or samples.

Composite elements must be editable, grouped, and semantic. They are not clip art. Each repeated primitive should encode a real object, category, feature, or data structure. Do not draw arbitrary chart glyphs with invented values. For detailed recipes, load references/composite-elements.md.

6. Classify the diagram archetype

Choose the closest archetype before layout:

• experimental workflow
• computational pipeline
• study/cohort flow
• method architecture
• mechanism schematic
• multi-omics/data-integration workflow
• graphical abstract
• hybrid multi-panel flowchart

Use one dominant archetype. Treat secondary archetypes as supporting modules, not competing layouts.

7. Load references only when needed

Use the references.on_demand table in manifest.yaml.

Operating Rules

• The scientific logic outranks style. Delete nodes and branches that do not carry a unique piece of evidence or procedural meaning.
• Prefer one clear reading path over a dense network. Use branches only for real alternatives, controls, or parallel assays.
• Draw the layout from a grid plan, not by eye. Nodes in the same semantic lane must share aligned x/y coordinates and dimensions.
• Avoid all-text diagrams. For each major scientific entity, ask whether it should be a grouped visual glyph made from draw.io primitives.
• Build composite glyphs from simple editable shapes, not raster icons. Group their child elements, use stable IDs, and keep a short label near the group.
• If a dedicated SVG has already been inserted for the same scientific object, treat that SVG as the visual glyph and do not also draw a duplicate primitive glyph unless the user explicitly asks for a hybrid editable reconstruction.
• Keep labels outside glyph interiors unless the label is an intentional symbol such as A/C/G/T on a base or a channel mark. Text must not cover primitive shapes.
• Treat connector routing as a first-class design object. No connector may pass through a text label, formula cell, node body, matrix tile, or module title.
• Minimize connector bends. Remove micro-jogs and redundant waypoints before final delivery.
• Render mathematical notation through draw.io MathJax syntax: set math="1" in the graph model and use \(...\) or $$...$$ delimiters.
• Use restrained journal style: white background, clean typography, subtle module grouping, consistent arrows, and low-saturation color families.
• Reclaim whitespace after replacing primitive glyphs with SVGs: recenter the SVG, move labels closer with safe padding, shrink stale placeholder regions, and reroute connectors with the fewest bends that remain valid.
• Make text editable and readable at final print size. Avoid rasterized labels, decorative shadows, gradients, and clip-art-like icons.
• Run scripts/qa_drawio.py on generated uncompressed .drawio files when Python is available, then fix all reported layout or source-integrity issues.
• For final or visually sensitive figures, export PNG/SVG previews with scripts/export_drawio_preview.py, inspect the preview, and iterate until inserted SVGs, formulas, labels, connectors, and canvas crop render correctly.
• For traced reference figures, run scripts/compare_drawio_reference.py against the exported PNG and use the diff overlays and worst-tile report to drive another edit pass before delivery.
• For close GPT/imagegen/raster tracing, run at least three strict export/compare/fix iterations and require a final strict-comparison pass before calling the figure complete. Failed strict metrics are not acceptable merely because the redraw is editable, vectorized, or formula-cleaned.
• Keep a private working trail private. Do not expose private paths, filenames, internal template names, or provenance unless the user explicitly asks.