Scientific Illustration Prompt Refinement: Before and After Examples

Circular four-step loop diagram for improving an AI figure prompt: draft, generate, compare, refine

The difference between a mediocre AI plate and a great one is rarely the model — it's two or three rounds of prompt refinement, done in the right order. This guide shows the anatomy of a strong figure prompt, real before/after rewrites, the one-change rule, and the patch-or-rewrite decision.

Practice on the presets · Full workflow · Errors and fixes

The anatomy of a figure prompt

A strong prompt is four slots, in order:

[SUBJECT + VIEW] → [PANELS] → [QUOTED LABELS] → [STYLE + TITLE]

Educational diagram breaking a strong figure prompt into four labeled parts

Subject + view — what it is and from where ("cutaway animal cell", "stratovolcano cross-section").
Panels — the named components of the layout ("whole specimen, one magnified inset, one comparison strip").
Quoted labels — the 3–6 terms in quotes, spelled exactly.
Style + title — the rendering register (styles guide) and the plate's printed title.

Vague prompts fail in slot 1 and 2; unreliable text fails when slot 3 is missing the quotes.

Before and after, four real rewrites

1. The vibe prompt

Before:

beautiful scientific illustration of a sunflower, very detailed, professional

After:

A mathematical-botany plate of the sunflower head: whole capitulum, a disc-floret
close-up, and an overlay spiral diagram showing Fibonacci parastichy families
counted and numbered in two directions. Label "CAPITULUM", "DISC FLORET", "RAY
FLORET", "PARASTICHY". Title: "COUNTING SPIRALS".

What changed: vibe words → subject + three named panels + four quoted labels. The before-prompt gets wallpaper; the after-prompt gets a figure. (This plate, and the workflow around it.)

2. The missing-constraint prompt

Before:

A labeled diagram of the human heart with all the parts

After:

An anatomically correct frontal cutaway of the human heart showing blood flow,
with oxygenated and deoxygenated paths, a small valve close-up and directional
arrows. Label "RIGHT ATRIUM", "RIGHT VENTRICLE", "LEFT ATRIUM", "LEFT
VENTRICLE", "AORTA", "PULMONARY ARTERY". Title: "THE HUMAN HEART".

What changed: "all the parts" is where mirrored anatomy and invented vessels come from — twelve vague parts become six exact ones. The label cap is a reliability feature, not a limitation.

3. The crowded-layout prompt

Before:

A figure showing everything inside a chloroplast, the whole photosynthesis
process, light reactions and Calvin cycle together, very comprehensive

After:

A Z-scheme figure of the light reactions: photosystem two splitting water and
sending electrons through plastoquinone, cytochrome, plastocyanin to photosystem
one, ferredoxin reducing NADP, and the proton gradient feeding ATP synthase.
Label "PSII", "PSI", "PLASTOQUINONE", "NADP", "ATP SYNTHASE".
Title: "Z SCHEME".

What changed: "everything" → one process, one path, five labels. Comprehensive prompts produce unreadable figures; scoped prompts produce the plate you actually reprint. Two figures beat one monster.

4. The error-correction prompt

Before (first output had the sill drawn diagonally):

A stratovolcano cutaway from crater to magma chamber: layered lava and ash strata,
central conduit, branching dike and horizontal sill... Label "CRATER", "CONDUIT",
"DIKE", "SILL", "MAGMA CHAMBER".

After (the fix):

A stratovolcano cutaway from crater to magma chamber: layered lava and ash strata,
central conduit, branching dike and horizontal sill — the sill strictly
horizontal, running between two strata layers, the dike cutting vertically across
strata... Label "CRATER", "CONDUIT", "DIKE", "SILL", "MAGMA CHAMBER".

What changed: one named constraint, appended. Not a reroll — a correction. This is the single highest-leverage habit in prompt refinement: fix errors by naming them, not by hoping.

The one-change rule

When iterating between two outputs you want to compare, change exactly one slot per regeneration — subject, panels, labels, or style. Otherwise you can't attribute the improvement, and you lose the working parts along with the broken ones. (The loop diagram at the top is this discipline: draft → generate → compare → refine.)

The loop:

Four-step prompt loop: draft, generate, compare, refine

Most plates converge in 2–3 cycles. If a prompt needs more than three targeted fixes, stop patching — the spec is wrong; rewrite from the four-slot anatomy.

Patch or rewrite?

Situation	Action
One label misspelled	Patch: shorten the label, regenerate
One structure misplaced	Patch: name the placement constraint
Layout unreadable	Rewrite: the panel list is wrong
Style inconsistent with the series	Patch: paste the series style clause verbatim
Three or more simultaneous fixes needed	Rewrite: you're patching a bad spec
The figure keeps almost working	Rewrite narrower: half the scope, twice the specificity

Scientific review checklist

Refinement is about the prompt; the checklist stays about the plate:

[ ] Are all labels spelled correctly?
[ ] Are all requested structures present?
[ ] Are positions and proportions reasonable against your reference?
[ ] Do arrows show the correct direction?
[ ] Are stages shown in the correct order?
[ ] Has it been checked against a trusted source?

FAQ

How many iterations until a plate converges?

Two to three for a well-specified prompt, one for lucky ones, and "never" for vague ones — the iteration count is a spec-quality meter. If you're on round five, the prompt is the bug: rewrite from the four-slot anatomy instead of patching. This is the patch-or-rewrite rule, and every experienced user eventually learns it the long way.

Should I save every prompt version?

Save the ones that produced keepers and the diffs that fixed real errors — that's your prompt library, and it compounds. The failed middles can go; what matters is the pair (before → after) that encodes the lesson. Ours became this article and the errors guide; a lab's version is a shared doc that new members inherit.

Why did my fix break something that was working?

Because the change wasn't isolated. Prompts are global constraints — an edit aimed at the sill re-weights the whole plate. That's the one-change rule: modify exactly one slot, regenerate, compare. When a fix must touch two things, do them in sequence, never in the same generation.

Is a longer prompt always better?

No — a more specified prompt is better, and length is a side effect of specification up to a point, then noise. The ceiling is roughly a screen: subject, panels, labels, style, and two or three named constraints. Past that, additions dilute the constraints that matter. The before/after pairs above all got shorter or equal while getting more specific.

Do prompts transfer between models?

The four-slot anatomy transfers; the constraint vocabulary partially. Subject-panels-labels-style is model-independent grammar, but which phrasings bite varies — one model's "nothing else" is another's ignored phrase. When switching, re-run your top five library prompts as a calibration set and note what changed; the library stays, the annotations gain a column.

Keep a prompt log

The cheapest habit in this whole discipline: a text file per project, one line per generation — date, prompt, what was wrong, what you changed. Ten seconds to write, and after a month it becomes a private textbook of what works on your subjects. It also makes figure provenance trivial: when a plate's accuracy is questioned months later, the log answers "which prompt, which fix, which reference" without archaeology (the verification artifact makes the same case from the reviewing side).

The prompt library pattern

After a month of refinement you own something valuable: a file of prompts that already work on your subjects, each encoding a lesson learned the long way. Structure it as pairs — the weak version, the fixed version, one line on why — exactly like this article. The library pays three ways: new figures start from a proven skeleton instead of a blank page (the anatomy); new team members inherit your iteration history instead of repeating it; and when a figure is questioned months later, the prompt that produced it is one lookup away. This site's version became the prompt packs — yours can stay a private doc. The format is the part that matters: before, after, lesson.

Reading someone else's prompt

A useful inverse skill: reading a prompt and predicting its failure. Scan for the missing slots — no named panels predicts crowding; no quoted labels predicts spelling errors; no style clause predicts drift; a subject without a view clause predicts the stylish three-quarter angle. You can practice on every prompt in our packs — each pairs with a verify note that names what actually needed checking, so the prediction gets graded. Prompt review becomes a code-review-like discipline: the four slots are the checklist, the verify note is the test, and the before/after pair is the diff.

Take your worst current prompt and rewrite it through the four slots on the scientific illustration presets. Related: errors and their fixes · label-count rule · style choices.