In my last post, How I model scientific claims before generating a diagram, I wrote about the thinking that happens before any image exists. This one is the hands-on part. I'll take a short text description of a CRISPR-Cas9 gene knockout, turn it into a graphical abstract, and then go through the output the way I'd review a figure from a co-author.
I built Scientific Figure Generator, so I'm using it here, but the prompting and review checklist apply to any image tool.
Why a knockout workflow is a good test
A gene knockout is a linear process most biologists can draw from memory, so when the generated figure gets something wrong, you notice quickly.
The five steps I wanted:
- Guide RNA design
- Cas9 delivery into cells
- Double-strand break
- NHEJ repair causing an indel
- Loss of protein expression
Writing the prompt
This is the exact prompt I typed:
Graphical abstract: CRISPR-Cas9 gene knockout workflow — guide RNA design, Cas9 delivery into cells, double-strand break, NHEJ repair causing indel, loss of protein expression. Flat illustration style, clean labels.
It's short, and that's on purpose. Three things are doing the work.
It names the figure type first. "Graphical abstract" at the start tells the model this is a single summary figure read left to right, not a results figure or a detailed pathway map. Leave it out and you tend to get something busier.
It lists the stages in order, separated by commas. I don't describe each panel. I give the sequence and let the order carry the logic. Each stage is short enough to work as a label. If I can't say a stage in about five words, I've probably packed two steps into one.
It names the style. The tool supports line art, 3D and flat styles. For a graphical abstract I usually pick flat, because it stays readable at thumbnail size. "Clean labels" nudges toward short, plain text.
I left out colors, panel counts and layout. The stage list already implies five panels in a row, and extra layout detail in a prompt this short usually hurts more than it helps.
You can also start from a sketch, a reference image, or an abstract or PDF, but here I wanted to see what a bare description produces.
What happens in the editor
With review mode on, the tool doesn't jump straight to an image. It first lays out a plan in the chat panel. In my run, the plan proposed a 21:9 wide format and offered 16:9 or 1:1 instead. It also noted that I'd given it a workflow directly rather than a paper abstract, so it built the plan from my stage list. After that comes a confirm step showing the job and its parameters (21:9, 1K) before anything is generated.
This is the cheapest place to catch a mistake. If the plan merged or dropped a stage, fixing it in text costs nothing. A 1K generation cost 10 credits in my test, so it's worth reading the plan before you confirm.
The output
Five panels, left to right:
Guide RNA design → Cas9 delivery → Double-strand break → NHEJ repair (indel) → Loss of protein expression
At first glance it's what I asked for. Here's what I actually check.
What to check before you use it
1. Scientific accuracy, one panel at a time
I go through each panel and ask, "Would a reviewer in this field object to this?"
- Guide RNA design: The panel shows a guide RNA with a stem-loop paired against a DNA helix. That's closer to "guide binds target" than to "design," but as an icon for the step it works. No PAM site is shown. For a general audience that's fine. For a methods-heavy paper I'd want it.
- Cas9 delivery: A green Cas9 shape carrying the guide is docked at the edge of a cell, and the nucleus is visible. It reads as delivery. It doesn't show the complex reaching the nucleus, where the cut actually happens. That's a simplification I'd accept in an abstract and question in a mechanism figure.
- Double-strand break: The helix is cleanly split into two pieces, and both strands are cut. Correct. Cas9 isn't shown at the cut site, which is a stylistic choice rather than an error.
- NHEJ repair (indel): The two ends are rejoined with a small yellow inserted segment. That shows an insertion. Indels can be deletions too, but the label says "indel," so the text covers the general case even though the picture shows one instance.
- Loss of protein expression: A dashed outline with an X through it. It's abstract, but the meaning is clear.
None of these are errors. They're simplifications, and whether they're acceptable depends on where the figure is going. Writing that judgment down is the step people skip.
2. Label correctness
Read every label letter by letter, since image models are known to garble text. Here all five labels match my stages exactly, and "NHEJ" is spelled right, which is the one I'd expect to break. There's a Check labels button in the toolbar, but I still read them myself.
Also check that each label sits under the right panel. A correctly spelled label under the wrong icon is worse than a typo, because nobody notices it.
3. Arrow logic
Each arrow should mean "leads to," and only that. In this figure there are four arrows, all pointing right, one between each pair of panels. No branches, no loops, no arrows pointing at nothing. For a linear knockout that's correct. If your process has a branch (say, NHEJ versus HDR), check that the fork actually splits from the right step and that both branches are labeled.
Exporting
Export gives you a PNG at 1K or 2K with no watermark. This one was generated at 1K. For anything going into a submission I'd pick 2K.
Starting from a template instead
If you don't want to start from a blank prompt, the template gallery is a quicker way in. There are 60 research templates, filterable by field (life sciences, chemistry and materials, physics and engineering, earth sciences) and by type (molecular and cellular, pathways and gene regulation, workflows and protocols, and so on).
Each card has an example prompt and a "Try changing" section, and you can use the template as a reference or edit a copy. The example prompts follow the same structure as above (figure type, ordered stages, style), so swapping in your own stages is easier than starting from scratch. There's a CRISPR Cytosine Base Editing template that's close to this example if you want to compare.
You can browse them here: scientificfiguregenerator.com/templates.
Cost
There are free daily credits, and paid plans start at $9.90/month.
The short version
- Put the figure type first, then the stages in order, then the style.
- Keep each stage short enough to be its own label.
- Read the plan before you confirm the generation.
- Review each panel for accuracy, every label for spelling and position, and every arrow for meaning.
- Decide which simplifications are fine for your audience, and write that decision down.
If you try this with your own workflow, tell me where it broke. Those are the cases I learn the most from.
Top comments (0)