A reference image and one short edit beat long prompts
Inkicon began as a case study: making a calligraphic icon family for a reading app. The most useful step wasn't a better prompt. It was a shorter one, applied to a sheet that was already close.
What happened
After several rounds of long, detailed prompts, the founder had a recognisable family that still wasn't right. They sent that sheet back to the image model with one line, spelling as typed:
make this more mediviel Calligraphy looking
The result was the direction they had been trying to describe all along.
Why it likely worked
The image already carried the structure: which icons, where, at what weight, and every decision made so far. The words only had to say what to change. A long prompt from a blank start has to specify all of that in text, and the model may honour some of it and drift on the rest.
Seen once more
We repeated the pattern later on a 16-icon set: the same short instruction on a full set. All 16 icons kept their place and meaning, no checks failed, and the pen character moved further than in any first generation — thin-to-thick strokes, crossed corners, diamond dots. It also made the lines thinner (0.54 px median at 16 px, against 0.83 px on the approved set), so a weight edit was the natural next step.
What it doesn't prove
Two observations are not a law. They don't show that short prompts always beat detailed briefs from a blank start, or that every edit keeps every icon intact. They do suggest a way of working: get a sheet that's close, then make small, focused changes and keep each version.
How Inkicon uses it
- Start from a style or your own reference image rather than a long description.
- Change the whole set with a short instruction, or redraw one icon while the rest stay put.
- Every version is kept, so you can step back.