Vivid Figures: an AI skill for research plots with fidelity guardrails

The gap Vivid Figures fills for research figures

Producing publication-quality figures usually means writing matplotlib or seaborn code by hand, tuning colors, spacing and layering until a chart reads well. Vivid Figures moves that work to an AI assistant. It is a skill: a folder of instructions and recipes that an assistant with file access, code execution and image viewing can load, then use to pick a chart type, generate it from your data, and check the result. The README's framing is that you do not need to know the name of the chart you want. You can hand it a CSV and ask for the plot that shows the difference between methods, and the skill selects a recipe. The audience is students and researchers preparing papers, math modeling reports and experiment writeups who want good defaults without hand-coding every figure.

From data to recipe to rendered figure

The skill follows the open Agent Skills specification: a root SKILL.md with YAML metadata and a Markdown body, and other files loaded by relative path as needed. The selection flow the README describes is staged. The assistant first reads the data and the goal, then consults explanation cards, then looks at candidate example images, and only then reads the full recipe code. This staged loading keeps the assistant from pulling every recipe into context at once. Color is handled by a single JSON palette file rather than per-recipe hardcoding, and project settings live under a .vivid/ directory. The README counts 143 figure recipes plus complete composite templates, with a catalog of 146 explanation cards and previews you can open in a browser, searchable by purpose, structure and tags. Output can be PNG for preview, PDF for layout, and the plotting source itself so you can keep editing.

Installing the Python dependencies and loading the skill

The skill needs an assistant that can read and write files, run Python and view images, plus Python 3.10 or newer with the repository's dependencies. Those dependencies are pinned in requirements.txt, which lists the scientific plotting stack: numpy, pandas, scipy, matplotlib, seaborn, scikit-learn and statsmodels among others. Install them into your environment with:

pip install -r requirements.txt

After the dependencies are in place, you load the whole skill folder into your assistant according to how that client imports or searches skills, then refer to it by name in your request. The README's install guide separates Windows from macOS and Linux steps and notes that plain data figures do not need an image-generation API key; only optional extras such as flowcharts or LaTeX technical diagrams pull in their own tools. You should see the assistant read your data, select a recipe, run the plot and produce a figures/ directory in your task folder.

Template fidelity: the rule that recipes may not be quietly simplified

The most interesting design choice is a constraint on the assistant, not a feature for the user. The skill requires the model to read the full recipe before drawing and to keep the template's gradients, transparency layering and key graphical elements, rather than collapsing them into flat colors or bare outlines to save code. Each explanation card lists the elements to preserve with source line numbers and the range of adaptation that is allowed. When you ask for edits, the guidance tells the model to prefer changes to position, spacing and size over redrawing. There is a check step that compares against the actual rendered image instead of trusting that a script ran without error. This is the project's answer to a common failure of AI plotting, where the model produces something plausible that has silently dropped the design intent of the template.

Where it depends on the model, and the license limit

Two limits are worth stating plainly. First, the README is explicit that these are execution requirements for the model, and the actual output still depends on whether the model follows them. The examples are not a guarantee of every run's result, and where the data does not support an element, the skill instructs the assistant to say so rather than invent, for instance not drawing a confidence band when there are no repeated trials. So the quality ceiling is set by the assistant you run it under. Second, the license is restrictive. The README states the skill is for personal, non-commercial use only, and forbids modification, derivative works, redistribution, resale or paid services without prior written permission. That rules out embedding it in a commercial product or an internal paid tool, and it constrains contribution in a way most open source figure libraries do not.

Vivid Figures versus hand-written matplotlib or a generic assistant

The plain alternative is writing matplotlib and seaborn yourself, or asking a general assistant to do it with no skill loaded. Hand-written code gives you full control and no license constraint, at the cost of the time it takes to design each figure and keep a consistent palette across a paper. A generic assistant with no recipes will produce a chart, but it has no shared palette, no fidelity rules and no catalog of vetted templates, so results drift between figures and often lose layering. Vivid Figures sits between them: it steers the assistant toward consistent, layered output with a fixed palette, in exchange for accepting its recipes and its non-commercial license. If you need commercial use or want to own the code outright, hand-writing the plots is the honest choice.

License and dependency footprint to check first

Before adopting, read the LICENSE file, because NOASSERTION on the repository means a custom, non-standard license rather than a recognized open source one, and the README's personal non-commercial restriction is the operative term. On maintenance, the last push was on 2026-09-16, with a tagged release v2026.09.15 shortly before, so the project has been receiving recent updates. The dependency footprint is not small: requirements.txt pulls in geopandas, PyMuPDF, CairoSVG and playwright alongside the core plotting libraries, so plan for a heavier environment than a bare matplotlib setup, and install only what your figures need where the README marks tools as optional.

Editorial conclusion

Adopt Vivid Figures if you are a student or researcher producing figures for personal or academic work and want consistent, layered plots without hand-coding each one. Do not adopt it for any commercial or redistributed product, because its license forbids that without written permission. Before you start, read the LICENSE file to confirm your use is allowed, and install requirements.txt into a dedicated Python 3.10 or newer environment, since the dependency set is large enough to be worth isolating.

Frequently asked questions

Official sources

1. Issues
2. README
3. Releases
4. yjz211/vivid-figures-skill on GitHub