If you've shipped a paper, you know the figure tax: you build a beautiful schematic, a reviewer asks you to move one box, and you spend an afternoon re-aligning everything by hand. Do that across three revisions and you've spent more time nudging vectors than running the analysis.
I started treating figures the way I treat code: as something that should be regenerable from a source, not hand-crafted once and patched forever. Here's the workflow that stuck.
1. Write the figure as text first
Before opening any drawing tool, I write the figure as a plain description in the repo:
ligand -> receptor (binding)
receptor -> kinase (activation)
kinase -> targetA, targetB (phosphorylation)
targetA -| receptor (inhibition, feedback)
This forces the logic to be explicit. If you can't write the relationships down cleanly, the figure isn't ready — and that's a science problem, not a drawing problem.
2. Generate a first draft instead of starting from a blank canvas
I feed that description into a tool that will create scientific figures with AI and get a laid-out first draft. The point isn't "AI does my figure." The point is the boring geometry — consistent boxes, even spacing, arrow routing — is done before I start, so my time goes into the parts that need judgment.
3. Keep the output editable and vector
The non-negotiable: the result has to be editable vector, not a flattened PNG. When the reviewer asks to move a box, I move a box. When a journal wants a different size, it scales without turning into mush.
4. Treat revisions as diffs
Because the source description lives in the repo, "revision 2" is a diff, not a rebuild. I change the text, regenerate, fix the few labels that matter, export. Twenty minutes instead of an afternoon.
What this doesn't solve
It won't make a bad figure good — encoding, color logic, and what you choose to show are still on you. And the first draft always needs cleanup; anyone promising one-click publication figures is selling something.
But moving the source of truth from "a fragile vector file" to "a text description plus a generator" changed figures from a dreaded last step into something I can iterate on as fast as I iterate on the writing.
How do you handle figure revisions? Still hand-editing in Illustrator, or have you found a more reproducible setup?
Top comments (0)