How to use the Academic Plotting agent skill for ML papers

What an agent skill is and what this one does

An agent skill is a folder of instructions and optional scripts that an AI agent loads when a task matches the skill's description. This skill is described as generating publication-quality figures for ML papers from research context.

It has two workflows. For architecture, system design, workflow and pipeline figures, the agent extracts components and relationships from the text and writes detailed prompts for Gemini (the skill names Gemini 3 Pro Image Preview). For data figures such as line charts, bar charts, scatter plots, heatmaps and ablations, it auto-selects a chart type from the data and writes matplotlib/seaborn code.

• Diagrams: the agent picks one of three visual styles (hand-drawn sketch, modern minimal, or illustrated technical) and applies it consistently across a paper.
• Charts: the agent chooses a chart type using rules such as line plot for step or time axes, grouped bars for comparing methods across benchmarks, and heatmaps for square matrices.
• Highlighting: the paper's own method is given a distinct colour so it stands out from baselines.
• Pie charts are discouraged in favour of stacked bars.

Install the skill (ZIP download or Claude apps)

Option 1: download the ZIP from the skill page on this directory and unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder so it applies only to that project.

Option 2: in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

The skill folder contains SKILL.md plus three reference files: references/data-visualization.md, references/diagram-generation.md and references/style-guide.md. Keep them together so the agent can read them.

Set up what the skill needs

The data workflow needs Python with matplotlib (3.8.0 or newer), seaborn (0.13.0 or newer) and numpy. The diagram workflow additionally needs google-genai (1.0.0 or newer) and a Google Gemini API key.

The skill's metadata lists these as its dependencies. If you only need charts, you can skip the Gemini key and the google-genai package.

Trigger the skill with a natural request

You do not need a special command. Describe the figure you want, paste the relevant paper text or your results, and say it is for a paper. The skill's description says it applies when creating any figure for a conference paper.

The agent reads your context first. It looks for components and connections in prose, or dimensions and metrics in numbers, before it chooses a figure type.

• "Make an architecture figure for the overview section from this method description: [paste text]"
• "Create a bar chart for this ablation table for our NeurIPS submission: [paste table]"
• "Plot training loss curves from this CSV for a conference paper"
• "Turn these benchmark results into a figure that highlights our method"

Example: from results table to chart

Suppose you give the agent three methods and two benchmarks: GPT-4 at MMLU 86.4 and HumanEval 67.0, your method at 88.1 and 71.2, and Llama-3 at 79.3 and 62.1. The skill's own example treats this as 3 methods by 2 benchmarks, selects a grouped bar chart, and highlights your method in coral.

The agent should then produce Python code you can run and inspect. Check the axis labels, the sort order and the colours before using the output.

Safety and what leaves your machine

The skill runs scripts and uses the network. Diagram generation sends your prompt to Google's Gemini service, so do not use the diagram workflow for confidential unpublished results or sensitive method details unless your policy allows that external call.

Data charts generated with matplotlib and seaborn are produced locally from the numbers you supply, but you should still review any script before running it.

• Diagram workflow: network call to Gemini, requires a Gemini API key.
• Chart workflow: runs Python plotting code on your machine.
• Check every generated label and component against your paper before submission.

Tips for better figures

Give the agent the exact terminology used in your paper, because the skill is designed to copy labels from the text. Pick one visual style for the whole paper, since the skill's guidance is that all figures should be consistent.

Keep diagrams to about eight top-level components. If there are more, ask the agent to group them into sections, as the skill suggests.

Export charts as vector PDF for LaTeX where possible; the style guide prefers PDF, and PNG for AI-generated diagrams, at 300 DPI minimum for raster images.