Type a few rough words. Get the image you actually meant.
“a funny cute cat”
Technically a cat. Nothing funny about it, no setting, no story. The model filled every gap with the most average choice.
A photorealistic full body shot of a fluffy, orange tabby kitten amusingly squished inside a clear glass jar labeled “COOKIE JAR - EMERGENCY SNACKS,” sitting on a rustic wooden bookshelf. The kitten has wide, wide-eyed expressions of slight confusion and a slightly tilted head, with delicate whiskers, tiny paws pressed against the glass, and soft fur details. Warm, soft natural window light creates subtle highlights and shadows across the scene, with sharp focus on the kitten’s face and bright catchlights in its eyes. The background features vintage books and a small potted succulent softly blurred with a smooth bokeh effect. The centered composition captures the hilarious yet adorable posture, with high fidelity rendering bringing out vivid textures in the glass, wood, and fur.
One of ten directions it hands back. The joke, the setting and the light are written in, so the model doesn’t have to guess.
“a funny cute cat”
Generated · flatWrite me a better image prompt for: a funny cute cat
**Create an image of:**
• a funny cute cat, photorealistic
• lighting: warm golden hour
• lens: 85mm, shallow depth of field
• mood: playful, candid
Reconstructed from the prompt logs. The same round trip, again and again.
It started with something in our prompt logs that didn’t look right.
Most people type the way you’d expect: a few words, no punctuation, “a funny cute cat.” But a lot of prompts had bullet points, headings, camera lenses, even lighting setups. Nobody writes like that in a small text box. Those prompts had been written somewhere else and pasted in.
So people were typing a few words, getting a flat image, going to another AI tool for better words, and coming back just to press generate.
They came to us for the picture and went somewhere else for the thinking.
I had six weeks to fix that.
Piktochart is a design tool for people who aren’t designers.
Bring the writing in-house, and the whole creative loop stays inside Piktochart.
If people step out for one job and come back, that isn’t a workflow. It’s a warning. Today they leave for the words. Tomorrow they stay in the other tab.
So stop asking people to write. Give them something to react to. My first answer to that was too expensive.
Obvious, and ten generations deep. Slower, pricier, every single time.
Recognition, done in text. Instant and free to render.
Your words, untouched.
Upgrades and generates. One tap, no reading.
Ten directions to choose from, then tune it.
Mood, setting, style, composition, narrative.
Scannable, so choosing feels like pointing.
Steer without going back to a blank box.
a funny cute cat
A photorealistic full body shot of a fluffy, orange tabby kitten amusingly squished inside a clear glass jar labeled “COOKIE JAR - EMERGENCY SNACKS,” sitting on a rustic wooden bookshelf. The kitten has wide, wide-eyed expressions of slight confusion and a slightly tilted head, with delicate whiskers, tiny paws pressed against the glass, and soft fur details. Warm, soft natural window light creates subtle highlights and shadows across the scene, with sharp focus on the kitten’s face and bright catchlights in its eyes. The background features vintage books and a small potted succulent softly blurred with a smooth bokeh effect. The centered composition captures the hilarious yet adorable posture, with high fidelity rendering bringing out vivid textures in the glass, wood, and fur.
Two prompts do all the work. The first turns a rough idea into ten directions. The second rewrites one of them when someone changes a setting. Every field they return has a place on screen.
titleCard headingvisual_attributesTags under each titlestyleSix-option picker staticdynamic_fieldsFraming and Lighting inferredsummaryDefault settings, preselectedfull_promptSent to the image modelGemini won on intent, cost and latency. Image-native models were worse at strict JSON, which was the whole contract. As a fallback, the same request went to ChatGPT instead.
Interrupting someone mid-creation breaks the rhythm. I lost that one.
V1 shipped as a capped experiment: the modal, every time. 96% completed, but the recordings still showed the interruption. V2 replaced it the next release with three modes — Off, On-Auto, On-Review.
Every AI call should cost credits. That’s how the rest of the product works.
Enhancement is a text call. It costs pennies, and this feature exists to stop people burning credits. Shipped free.
Piktochart announced it on the blog, documented it in the help centre, and posted two reels on Instagram. That’s the promotion behind April’s number on the next slide.
people tried it by early June
got their ten directions back
of them generated an image with one
April was launch month, with promotion on Piktochart’s channels. May had none and still held at 95%. March is hatched because it launched late in the month.
The last bar is the fallback, not a drop-off. Nobody hit a dead end.
| Month | Unique users |
|---|---|
| Mar 2026 | 362 |
| Apr 2026 | 16,434 |
| May 2026 | 15,560 |
| Funnel step | Users |
| Enhance triggered | 34,170 (100%) |
| Enhanced prompt generated | 32,850 (96.1%) |
| Enhanced prompt used | 24,810 (72.6%) |
| Fell back to original prompt | 668 (2.0%) |
Every figure here is against the 34,170 people who turned the enhancer on inside the AI image generator — not Piktochart’s 34M registered accounts, which this feature never sat in front of.
Cut on cost.
Cut on principle.
Cut for the deadline. It comes back.
Cut for speed. This is the one I’d put back.
“The logs told me what people were doing. Interviews would have told me why. I had six weeks, so I bought the what.”
The call I’d take backSame question behind every one of them: where is the user doing work the product should be doing for them?
Turns a few rough words into ten directions for the image generator.
Generates video from a text prompt or an image, with a choice of model.
Generates a logo from a short description, in illustrated, engraved or line-art style.
Rewrites any text in a design: shorter, longer, clearer, corrected, or in a different tone.
Increase spacing between text blocks and images
Scores a design out of ten and lists what to fix, by layout, spacing, colour, readability and hierarchy.
Turns the text and visuals in a flat image into fully editable layers.
It was giving people something to react to.
Most of this feature was designed from a terrace in Sa Pa, Vietnam. That’s the review modal on the screen.
Click any card to jump straight to it