Scientific Illustration

Paragraph in. Figure out.

An open-source skill for AI coding agents that turns a paragraph of your paper into a publication-ready figure, 3D image, lesson summary or short video — written as editable code, rendered with free medical-grade art.

Website · Full video · Examples · Changelog

---

What you get

You ask for…	You get	Edit it later?
A scientific figure / graphical abstract	SVG · PNG · PDF + a browser editor	✅ drag, recolour, retype
A plot of your data (FTIR, XRD, TGA, tensile, release…)	Multi-panel figure, vector text	✅
A flowchart, PRISMA or mind map from text	SVG · PNG	✅
A 3D model or 3D image	Interactive 3D page + transparent PNG (up to 4×), or a photoreal Blender Cycles render	✅ in the page / in the spec
A lesson summary	Editable A4 / slide card · PNG · PDF	✅ type directly in it
A process simulation	Interactive HTML with sliders	✅
A short video	MP4 with an original soundtrack (Remotion)	✅ change the JSON, re-render

Researches first: give it just a topic ("a figure on ferroptosis") and it pulls recent, highly cited papers from OpenAlex, Europe PMC and Wikipedia (free, no key), then draws only what the sources support and cites them as [n]. Two levels: a simple version (one message, 3–6 elements, for slides and teaching) or a detailed one (multi-panel, values, references, for papers) — or both from the same spec. Two styles: Academic (clean, publication-ready) or Pixel art (game look, for social media and teaching) — the skill asks which you want. You don't need to name the format. The skill reads what you attach, where it will be used and how you ask, then picks and tells you its choice.

Input → output

You write: “Show how an organosilane dispersant keeps silver nanoparticles evenly spread in a wound dressing — with vs without, and the failure modes when dispersion is poor.”

The agent wrote this spec; change any word or colour in it and render again.

Gallery — code-drawn (Route B)

How it works

Every job starts with a choice of two routes:

• Route A, with AI: an image AI you choose (you sign in yourself in the agent's browser; Gemini is the tested fallback) draws only the scene. The agent adds every label, draws every molecule from SMILES, checks the image for errors in crops and writes the AI disclosure. No AI micrographs, spectra or data plots.
• Route B, without AI: everything is drawn by code, fully reproducible, no disclosure needed.

Either way the agent first asks a few multiple-choice questions about your project, searches the literature and shows you the references, and remembers your feedback (client_profile.md in your project folder) for the next request.

1. The agent writes a short JSON spec (a few thousand tokens) instead of drawing pixels.
2. Python / Node scripts render it. Same spec → same result, every time.
3. The agent looks at the result, fixes overlaps, and hands you the output together with its spec.

Install

Requirements: Windows, macOS or Linux · Python 3.10+ · Node 18+ (for 3D and video) · ~300 MB for video tools, ~1 GB optional for the art libraries. Optional: Blender for photoreal 3D (tested with 5.1; GPU recommended), and Chrome or Edge for export.

git clone https://github.com/charaflokman/scientific-illustration ~/.claude/skills/scientific-illustration
python ~/.claude/skills/scientific-illustration/scripts/setup_check.py --first-run

setup_check.py lists what is ready and what is missing, with the install command for each piece; the agent runs it on first use and asks before installing anything. Python tools for specialised 3D (ASE, OVITO, PyMOL, cellPACK, TPMS scaffolds) are optional and live in their own virtual environments.

First test (10 minutes)

1. Open your agent in an empty project folder and ask, in any language: "Make a graphical abstract of chitosan films with ZnO nanoparticles for food packaging."
2. Answer the route question (A or B) and the short project questions (e.g. 1a 2a 3b 4b …).
3. Check the references it shows, then open the .editor.html file it gives you: double-click any label to retype it, drag elements, and save the SVG.
4. Tell it what you like or dislike ("too simple", "the gel looks like glass"): it records the feedback and redraws.
5. Try your own data: "Plot my FTIR and TGA CSVs as panels (a) and (b) with a T5% table."

Other agents: clone into ~/.codex/skills/, ~/.gemini/skills/ or .agents/skills/, or point your agent at AGENTS.md. Works with Claude Code, Codex CLI, Gemini CLI, Cursor, Copilot, Cline, OpenCode and others, with any model.

Then ask in plain words, in any language:

"Make a graphical abstract from this methods section" · "Plot my FTIR and TGA CSVs as panels (a) and (b)" · "Summarise this lesson as one card" — in any language

Free art libraries (about 23,000 drawings)

python scripts/library.py fetch downloads them once (--only reactome phylopic for a subset); library.py search fibroblast searches all of them. Credits are added to the figure automatically.

This is the beginning of the README. Read the rest at the original. Read the full README ↗