keros68/xiaoyu-skill/abstract-fig

Abstract-Fig

Turn manuscript content into a figure whose scientific message, module relationships, and visual hierarchy are clear at its intended publication size. Keep the source editable. Draw.io is the default; follow an explicit request or an established project route for script-generated SVG (Inkscape), PowerPoint, or another suitable editable format.

Working contract

• Preserve the previous figure and edit a versioned copy or keep a recoverable backup.
• When the user has edited a version by hand, treat that file as the base: change identified elements only and keep the user's font, proportions, alignment, and wording.
• Use manuscript evidence and actual project materials. Preserve scientific meaning, numerical values, uncertainty, and the distinction between observations and schematics.
• Keep text, connectors, and structural objects editable. Embed reused images; explain that pixels within a raster remain raster content.
• Deliver the editable source and a rendered preview. Export submission PDF/SVG/PNG when requested or part of the established manuscript workflow.
• Image generation is optional. Neither image count nor use of a particular tool is a completion criterion.
• Apply user instructions and existing design decisions before defaults. A local adjustment does not require redesigning the whole figure.
• Check the journal's policy before using generated imagery. Where generative-AI imagery is not accepted, a draft with generated elements is a design sketch and those elements must be replaced before submission.

Workflow

1. Define the communication task

Identify the figure's role, the research question it supports, and what readers should understand from it. Consult figure-types.md when the role is unclear. Distinguish a method framework from a graphical abstract or a results synthesis; they need different information density. For a graphical abstract, read graphical-abstract.md before drawing.

Read the relevant manuscript passages, captions, existing figure, and available assets. When the user keeps a reference collection or supplies examples, inspect them first and identify their layout family; follow that family unless the content cannot fit it. For each proposed module, identify its purpose, essential content, supporting source, and connection to other modules. Omit implementation inventories and details better carried by the caption. Do not reproduce the section list as a row of boxes.

2. Choose the content and module arrangement

Use style-decision-gate.md for content planning, reference comparison, and module combinations. Define sequence, comparison, shared input, parallel analysis, convergence, or validation relationships before placing boxes.

Allocate area according to importance and the space needed to read the evidence. Align comparable objects, but do not force unrelated modules to equal sizes. Choose canvas proportions and density from the content and target display width, not a fixed panel count or aspect ratio.

For a new or substantially redesigned figure, briefly state the content, reading path, grouping, and intended visual treatment. Continue when the request, references, or existing decisions provide enough direction. Ask only about missing information or materially different interpretations; do not impose an approval menu.

3. Select representations and sources

Use as many elements as the content requires, including zero raster images. Never generate or retouch an illustration to stand in for measured imagery, a geographic boundary, an experimental photograph, or a quantitative result. Label illustrative examples and hypothetical states when readers could mistake them for observations.

Only when generating illustrations, read image2-element-workflow.md. Missing image2 does not block vector construction or use of existing assets. If an essential source is unavailable, identify that specific gap without inventing a replacement observation.

4. Build the figure

Use boxed-manuscript-style.md for hierarchy, typography, lines, and legends. For draw.io assembly, read drawio-element-workflow.md; for script-generated SVG edited in Inkscape, read svg-inkscape-workflow.md. For proposal or project roadmaps, consult research-roadmap.md.

Shorten text before reducing font size. Put detail in the caption where appropriate. Preserve Chinese, Greek letters, subscripts, superscripts, units, and other scientific notation; write subscripts as real formatting, not Unicode subscript characters. Calibrate wording and arrows to the evidence; a process relationship is not automatically a demonstrated causal mechanism.

5. Inspect the actual deliverable

Follow qa-checklist.md. Render the editable file in its intended editor/exporter when available, then inspect the expected publication width. A separately generated preview can aid development but does not prove that the editable file renders identically. State any unverified rendering limitation.

For draw.io, use python scripts/inspect_drawio_images.py <figure.drawio> to check image embedding and identify images needing review. This supports vector-only, uncompressed, and compressed files. Large image warnings require visual judgment; a large map is not automatically a flattened diagram. Use --min-images only when the particular design requires a known number of images.

6. Hand off

Link the editable source and preview; briefly report the substantive change, validation, and any material limitation. Identify reused or generated assets where relevant. Avoid reporting image counts as evidence of scientific or visual quality.

For draw.io, explain once when useful: open https://app.diagrams.net/ and drag in the .drawio file. Text and structural objects remain editable; embedded images can be moved, resized, or replaced. Other formats should have equivalent handoff instructions.