Description
Figures are usually the part of a paper or modeling report that eats the most time. Matplotlib defaults look out of place the moment they land in a manuscript, and tuning font sizes, palettes and legend placement can burn an afternoon without landing anywhere good. Handing the job to an AI is not obviously better — it tends to grab a bar chart and call it done, picks different colors every run, and nothing in the set looks related. Vivid Figures takes another route: it hands the model 141 chart recipes together with their full source drafts, so the AI adapts a vetted template to your data instead of improvising one.
You give it the data and say what you want the figure to show. It reads the fields, browses the spec cards, compares real preview images, pulls the complete recipe for the chart it settled on, renders, then looks at the output and fixes what is off. You do not need to know the name of the chart you want — "compare these methods and pick something that makes the differences visible" is enough — though you can also ask for a ridgeline, raincloud or heatmap by name. What comes back is a PNG, a PDF, and plotting source you can keep editing.
141 chart recipes: from grouped bars and multi-series lines to ridgelines, Sankey diagrams, survival curves, 3D surfaces, parallel coordinates, SHAP summaries and calibration plots — each with a working recipe and a real rendered preview.
Seven palettes: Coral Teal, Olive Apricot, Blue Pink, Blue Sky, Soft Forest, Pastel Girl and Ocean Breeze, each with a maintained categorical order and continuous ramp, plus custom colors if you want them. Olive Apricot is the default, and follow-up figures in the same session reuse whatever you already picked.
Chart choice before rendering: the model follows a fixed path — understand the data, read the spec cards, look at candidate renders, then load the full recipe — rather than guessing a chart type from column names and turning a distribution into a bar chart.
Template fidelity rules: every spec card lists what must survive, with source line numbers, and what may be adapted. Gradients, transparency layering and key graphical elements cannot be flattened into solid fills or bare outlines just to shorten the code; edits favor position, spacing and size.
Browsable visual catalog: the repo ships catalog/index.html, a gallery of all 141 template cards and previews that opens in any browser, searchable by purpose, structure and data requirements and filterable by tag — handy when you would rather pick the chart yourself.
Three deliverables: PNG for quick review, PDF for typesetting, and the plotting source for reproduction and further edits, normally written to a figures/ folder in your task directory.
Revisions keep the look: "move the legend up and bump the font size" continues from the existing figure, keeping the template and palette already chosen instead of drifting into a different style.
No invented data: when the data cannot support an element, the model is told to explain the adjustment rather than fake it — no confidence band conjured out of a single run.
Open skill format: built on the Agent Skills specification, with a root SKILL.md using standard YAML metadata and Markdown body, so any assistant that supports the spec and can read files and run Python can load the folder.
Requirements: Python 3.10+ and the repo's dependencies, Git for cloning and Bash for the bundled check scripts (Git Bash works on Windows). Flowcharts and LaTeX technical figures need extra tooling; ordinary data plots do not.
License note: the author restricts it to personal, non-commercial use — modification, derivative works, redistribution and paid services all require prior written permission, so read the repo's LICENSE before any commercial use.
You give it the data and say what you want the figure to show. It reads the fields, browses the spec cards, compares real preview images, pulls the complete recipe for the chart it settled on, renders, then looks at the output and fixes what is off. You do not need to know the name of the chart you want — "compare these methods and pick something that makes the differences visible" is enough — though you can also ask for a ridgeline, raincloud or heatmap by name. What comes back is a PNG, a PDF, and plotting source you can keep editing.
Features
141 chart recipes: from grouped bars and multi-series lines to ridgelines, Sankey diagrams, survival curves, 3D surfaces, parallel coordinates, SHAP summaries and calibration plots — each with a working recipe and a real rendered preview.
Seven palettes: Coral Teal, Olive Apricot, Blue Pink, Blue Sky, Soft Forest, Pastel Girl and Ocean Breeze, each with a maintained categorical order and continuous ramp, plus custom colors if you want them. Olive Apricot is the default, and follow-up figures in the same session reuse whatever you already picked.
Chart choice before rendering: the model follows a fixed path — understand the data, read the spec cards, look at candidate renders, then load the full recipe — rather than guessing a chart type from column names and turning a distribution into a bar chart.
Template fidelity rules: every spec card lists what must survive, with source line numbers, and what may be adapted. Gradients, transparency layering and key graphical elements cannot be flattened into solid fills or bare outlines just to shorten the code; edits favor position, spacing and size.
Browsable visual catalog: the repo ships catalog/index.html, a gallery of all 141 template cards and previews that opens in any browser, searchable by purpose, structure and data requirements and filterable by tag — handy when you would rather pick the chart yourself.
Three deliverables: PNG for quick review, PDF for typesetting, and the plotting source for reproduction and further edits, normally written to a figures/ folder in your task directory.
Revisions keep the look: "move the legend up and bump the font size" continues from the existing figure, keeping the template and palette already chosen instead of drifting into a different style.
No invented data: when the data cannot support an element, the model is told to explain the adjustment rather than fake it — no confidence band conjured out of a single run.
Open skill format: built on the Agent Skills specification, with a root SKILL.md using standard YAML metadata and Markdown body, so any assistant that supports the spec and can read files and run Python can load the folder.
Requirements: Python 3.10+ and the repo's dependencies, Git for cloning and Bash for the bundled check scripts (Git Bash works on Windows). Flowcharts and LaTeX technical figures need extra tooling; ordinary data plots do not.
License note: the author restricts it to personal, non-commercial use — modification, derivative works, redistribution and paid services all require prior written permission, so read the repo's LICENSE before any commercial use.
