AI Prompt Enhancer
Piktochart could make the image.
Users couldn't ask for it.
Piktochart is an AI design platform for non-designers. 34M+ users worldwide have made 220M+ visuals: infographics, presentations, reports, social graphics, and AI images.
The generator was fine. The prompting wasn't. Most users aren't prompt writers. They'd type something rough, get a flat image, then go find a prompt tool somewhere else and paste the result back.
An in-app image prompt enhancer that gets users a better image on the first try, without leaving Piktochart
Users were voting with their feet
Generate in Piktochart
Type a prompt, hit generate.
Hit friction
Mediocre result, unclear why.
Leave Piktochart
Use another tool's prompt enhancer.
Paste back to generate
Return only for the image.
Not a feature request.
A retention leak.
Nobody asked for a prompt tool. They wanted one less step. They knew their prompts were weak, they just couldn't fix them alone, and guessing felt worse than switching tabs.
The realization
- The pattern was repeatableGenerate → friction → leave → paste back
- The fix was structuralBring the prompting in-house
- The prize was the whole loopIdea, refine, generate, without leaving
Keep them here.
Give them a better image.
For users
- Skip the guesswork
- One toggle, several clear directions
- Better images on the first try
- No reason to open another tab
For business
- Keep the users we were losing
- Make each generation worth more
- Own the whole workflow, not just the output
- Something competitors don't have
One toggle.
Three modes.
Better image.
Uses the original prompt without changes.
Improves the prompt and generates straight away.
For people who want more control. It reads what the prompt is going for and offers 10 rewrites. Any of them can be pushed further:
From a vague idea to a usable prompt
It reads what the prompt is going for, then writes 10 versions of it. Each one gets a short title and a few tags.
Mood Focused, Action & Drama, Artistic & Minimal, and so on. The full text is one tap away.
Six styles to pick from, plus Framing and Lighting. Change any of them and the prompt rewrites itself.
One click. No prompt writing needed.
Skips all of this. It picks the closest match and generates straight away.
Two model calls, one strict contract
The model doesn't return prose. It returns an interface. Every field is something you can see, so I wrote the schema and designed the screen at the same time.
User idea in. Ten suggestions out, each with a title, tags, and a generation-ready prompt.
The chosen prompt plus the user's Style, Framing and Lighting picks. One refined prompt out.
Every edit links back to the one before it, so users can walk back through their changes.
titleCard headingvisual_attributesTags under each titlestyleSix-option picker staticdynamic_fieldsFraming and Lighting inferredfull_promptSent to the image modelHow do you judge an image
you can't see yet?
Two ways to solve it. I bet on the cheaper one.
Preview image per suggestion
Slower, pricier, more API calls. All for a problem users might not have.
Trust structured text
Short titles, tags for the visual bits, the full prompt on tap. Cheap and fast, and people got it straight away.
They found it,
and they finished.
people tried it in the first month
turned it on and finished a generation
better results, so fewer second attempts
Three assumptions that turned out right
Text-only was enough
People read visual direction straight from text. No previews, no image cost, and it shipped faster than planned.
The model handled rough prompts
Tight instructions meant even a lazy prompt came back with ten usable directions. Nobody had to write well to get something good.
It kept users in
One toggle, ten suggestions, one click. People stopped opening other tabs. That was the whole point.
Promoted in public, on Piktochart's own channels
Piktochart promoted the feature on its own Instagram channel:
What I'd carry into the next AI feature
Write the instructions like a spec
The tighter I wrote the rules on structure, tone and wording, the more predictable the model got. Most of the design work was in the prompt.
"One-time use" can be a win
Repeat use dropped because the first try worked. A number only means something next to the goal.
Text-first UX is underrated
Structured text beat image previews on speed and cost, and people understood it just as well.
What's next
- Design ReviewAI feedback on existing work
- AI CopywriterText suggestions, not just images
- TemplatizerTemplates matched to intent
- Deeper customization UXUsers under-explored it
Thanks for reading.
Happy to go deeper on any of it: the decisions, the prompt work, or the numbers.
