菜单
Guidance for adapting paper figures, screenshots, and reference images to user data in ggai.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | ggai-reference-figure |
| description | Guidance for adapting paper figures, screenshots, and reference images to user data in ggai. |
| when_to_use | Use when the user says to draw a plot like a reference image, article figure, screenshot, paper figure, example chart, or style sample. |
Turn reference-figure requests into honest, data-faithful ggplot work. The reference may be a local image path, article path, URL, screenshot, paper figure, method description, or a phrase such as "use my data to draw this kind of plot".
Do not require fixed arguments such as reference_image= or reference_mode=.
Infer the reference role from the user's natural-language request and available
context.
Classify the reference before acting:
One request can combine several roles. Prefer this priority:
For structure references, preserve the intent rather than the pixels:
For style references, preserve only styling:
When adapting a reference, do not copy its density blindly:
Commit when the plot:
Declare a blocker only when:
Generate a single biomedical entity as a reusable cutout asset (transparent alpha background, no labels) suitable for compositing into diagrams. Use when the goal is to produce one isolated subject — a cell, vessel, tissue cutaway, signaling cloud — that will later be placed into a scene, rather than rendering a complete figure.
Produce a schematic, illustration, biomedical cartoon, or pure conceptual figure without underlying data. Calls an image model directly with a structured prompt, generates one or more candidates, scores them on basic visual proxies, and saves the best. Use when the goal is "draw / illustrate / diagram / sketch" and no data frame is in scope.
Decide which R visualization engine to use for a figure: ggplot2, grid/grob, base graphics, ComplexHeatmap, circlize, ggraph/DiagrammeR, htmlwidget (plotly/leaflet), or composite (patchwork/cowplot/aplot). Load this skill when the user names a non-ggplot library, when the figure shape suggests a specialized library, or when ggplot would require contortions to reach the intended result.
Produce interactive web visualizations using `plotly`, `leaflet`, `DT`, `networkD3`, or any other `htmlwidgets`-based library. Default output is a self-contained HTML file (always works). Static PNG export is supported when the optional `webshot2` package and a headless Chrome (via `chromote`) are available.
Compose multiple plots into a single figure using `patchwork`, `cowplot`, or `aplot`. Use whenever the user wants more than one panel — labeled multi-panel figures, side-by-side comparisons, A/B/C panels with shared legend, plots stacked over a metadata strip, or any layout that exceeds what a single ggplot's `facet_*` can express.
Lift the visual quality of an existing data figure by redrawing it with an image model that uses the original ggplot as a hard constraint. Preserves data semantics (positions, groups, scales, labels) while improving typography, composition, and surface polish. Use when the user has a working figure but wants it to look like a publication, cover, or polished talk slide.
