research-figure-studio

name

research-figure-studio

description

Use when the user intends to create, redesign, polish, audit, or submit social-science research figures, paper figures, scientific illustrations, mechanism diagrams, conceptual frameworks, workflow diagrams, research-design figures, result figures, multi-panel figures, or image2-assisted academic visuals. Automatically use this skill for requests to draw figures for papers, reports, manuscripts, top-journal review, or research presentations.

Research Figure Studio

This skill is the default workflow for social-science research figures and paper-ready scientific illustrations.

Trigger

Use this skill whenever the user asks to:

• 画论文图、科研图、论文图表、学术图表、机制图、概念图、流程图、研究设计图、结果图、多面板图、顶刊风格图。
• 出图、制图、配图、图表绘制、图表美化、结果可视化，尤其是论文、报告、答辩和审稿场景。
• improve an existing figure for manuscript review, submission, report, defense, or academic presentation.
• use image2, image generation, diagram generation, SVG, PPT, HTML, Python, R, or other tools to make a research figure.

Do not treat this as a generic poster, marketing graphic, dashboard, or app UI task.

Required References

Before producing a serious figure, read only the relevant parts:

• references/social-science-top-journal-figure-style.md
• references/image2-figure-contract.md
• <VELA_RUNTIME_ROOT>/skills/catalog/scientific_figure_workflow.json
• <VELA_RUNTIME_ROOT>/skills/catalog/figure_style_presets.json
• <VELA_RUNTIME_ROOT>/skills/catalog/publication_style_rules.json

Core Rule

Research figures must be argument-bearing evidence objects. They must show a claim, a mechanism, a design, a comparison, a model, a data result, or a workflow that the surrounding manuscript actually needs.

Scientific Figure Standards

Use scientific_figure_workflow.json as the shared contract for figure production standards.

For conceptual, mechanism, research-design, workflow, and image2-first figures:

• Lock the exact structure, text, arrows, panel count, palette, and unsupported-claim prohibitions before rendering.
• Use social_science_nature_red_blue_rainbow as the default formal research figure preset unless the user or target journal requires another preset.
• Apply the typography contract unless the target venue requires another house style: serif fonts; Chinese fallback SimSun, Songti SC, Source Han Serif SC, Noto Serif CJK SC; English fallback Times New Roman, Liberation Serif, DejaVu Serif; base 10 pt; captions 9 pt; panel labels 10 pt.
• Keep figure text short enough to remain readable at single-column or double-column manuscript sizes.
• Prefer PDF/SVG/PPT/editable source for final use, with PNG preview as needed.

For empirical data panels inside a mixed figure set:

• Delegate the numeric chart/table panel to figure-table-studio.
• Require data health, process data, statistical justification, caption, and export records before the panel is accepted.
• Do not use image2 to generate or repair empirical numbers, p-values, sample sizes, axes, error bars, or source notes.

Renderer Policy

image2 is allowed and should not be refused merely because the target is a research or paper figure.

Choose the renderer by figure type:

• Exact data plots, regression outputs, event-study charts, coefficient plots, descriptive plots, maps from real data: use deterministic code first. Python, R, Stata export, SVG, PDF, or PPT are preferred. image2 may help with visual layout only if all numbers, labels, and geometry remain controlled.
• Conceptual frameworks, mechanism diagrams, process diagrams, research-design diagrams, typology figures, visual abstracts, theory maps, and non-numeric scientific illustrations: image2 is a valid first-class renderer after the structure, labels, arrows, and style are locked.
• Text-heavy figures: prefer SVG, HTML, PPT, or code-rendered diagrams, or post-edit image2 output with deterministic text. Do not accept unreadable or invented text.

Never let image2 invent empirical values, axes, statistical significance, sample size, source notes, equations, citations, or causal claims.

One-Pass Workflow

1. Detect figure intent and select this skill automatically.
2. Build a figure brief before rendering:
   • figure purpose
   • manuscript claim or research function
   • figure type
   • data/source basis, or state that it is conceptual
   • panel plan
   • exact text labels
   • arrow and relationship definitions
   • style target
   • style preset id
   • renderer choice
   • typography and size defaults
   • title/caption outside-image plan
   • overlap risk checks
   • risk checks
3. If the user supplied data or a draft figure, preserve the substantive content first.
4. If the user did not supply enough detail for a conceptual figure, make conservative assumptions and label them in the brief. Do not block unless the figure would imply unsupported empirical evidence.
5. Render the figure.
6. Audit the output before final delivery.

Social-Science Top-Journal Style

Default style:

• white background
• red-blue anchored Nature-style rainbow color
• black-and-white readable
• color-blind friendly
• consistent typography
• consistent line width
• direct labels when possible
• legends under or inside the figure, not floating on the side unless a target journal requires it
• no decorative gradients, glassmorphism, 3D effects, bokeh, stock-photo ambience, or marketing layout
• no formal figure title or long caption inside the image; title, caption, source, sample, model note and script path belong in manuscript text, figure notes or delivery notes

Multi-panel figures:

• every panel answers a distinct question
• panels are equal or intentionally proportional in size
• panel labels are short and meaningful, such as (a) Diffusion threshold
• shared legend appears once
• panel sequence should match the evidence chain

For social-science manuscripts, prefer this evidence order:

1. descriptive fact or setting
2. main estimate or central comparison
3. mechanism, heterogeneity, or process
4. robustness, sensitivity, or boundary

Style Presets

Use figure_style_presets.json before rendering:

• social_science_nature_red_blue_rainbow: default for formal mechanism, concept, research-design, workflow, typology and mixed overview figures. It uses white or very light gray background, high whitespace, fine lines, black text, red-blue semantic anchors, and controlled cyan, amber, orange and indigo rainbow accents for groups, gradients, maps and multi-series figures.
• nature_empirical_red_blue_rainbow: use when this skill coordinates a mixed figure set that contains empirical panels; delegate those panels to figure-table-studio.
• minimal_review_ready_red_blue: use for submission, reviewer response and final manuscript figures when review safety is more important than visual emphasis.
• presentation_premium_red_blue_rainbow: use for defense, report and presentation figures when screen readability and visual hierarchy matter, while preserving evidence boundaries.

Do not describe the target only as "Nature style", "top-journal style" or "make it premium". Pick or inherit one preset and record it in the figure brief.

Overlap And Caption Boundaries

Formal figures must pass these checks before delivery:

• Text must not overlap modules, arrows, panel labels, legends, or other text.
• Legends must not overlap lines, points, confidence bands, bars, arrows, or labels.
• Arrowheads must not touch text or pass through label boxes.
• Panel labels must not collide with axes, module titles, legends, or content.
• Figure title and long caption must stay outside the image.
• Data source, sample, model note, uncertainty note and script path must stay outside the image unless the target venue explicitly requires a short in-figure source note.

image2 Prompt Contract

When using image2, the prompt must include:

• "white-background social-science journal figure"
• "red-blue anchored Nature-style rainbow color palette"
• the selected style preset id, usually social_science_nature_red_blue_rainbow
• exact panel count and layout
• exact visible text labels
• exact arrows or relationships
• controlled red-blue rainbow palette
• "no figure title inside the image"
• "no long caption inside the image"
• "no overlapping text, legend, arrows, points or lines"
• no fake data, no invented axis values, no invented citations
• no decorative background, no 3D, no marketing poster style
• "all text must be readable and spelled exactly as provided"

If the figure has many Chinese labels, use image2 for visual structure only when text can be verified or repaired afterward. Otherwise render with HTML/SVG/PPT.

Architecture Figure Adapter

For method architecture, mechanism, workflow, or conceptual framework figures, the local workflow absorbs the strongest constraints from Leey21/awesome-ai-research-writing:

• Understand the abstract, method, theory, or research design first; do not draw a generic pipeline.
• Highlight the core novelty or central mechanism, but only if it is actually supported by the manuscript.
• Use a white-background flat vector academic style with clean lines, restrained pastel tones, and no photorealism, decorative 3D, poster style, or unreadable text.
• Group related components logically and make arrow directions reflect real data flow, causal logic, process order, or conceptual dependence.
• Keep visible text short and exact. Long explanations belong in the caption or surrounding manuscript, not inside the figure.

Required Audit

Before saying the figure is done, check:

• Does the figure support a specific manuscript claim or research function?
• Are all labels readable and free of hallucinated text?
• Are arrows and module relations correct?
• Are there any overlaps among text, legends, lines, points, arrows, panels, and labels?
• Is any data, p-value, axis, sample size, DOI, or source note invented?
• Can it still be understood in grayscale?
• Do typography and font sizes follow the locked contract or a stated target-journal override?
• Is the caption able to explain unit, sample, uncertainty, source, and script path when applicable?
• Are formal figure titles and long captions kept outside the image?
• Are editable or high-resolution outputs available when needed?

If any audit item fails, revise before final delivery.

Output

For manuscript work, deliver:

• final figure file
• editable source when possible
• figure brief
• caption or figure note
• audit checklist

For quick discussion figures, at minimum deliver the image and a concise note of assumptions.