You Are Not Getting Jarvis. Here Is How to Work With AI
Introduction: The Donkey From Shrek
Yuliia Chumak started using AI expecting Jarvis — smart, loyal, always one step ahead. What she got was the donkey from Shrek: talks a lot, tries hard, occasionally gets things right, and mostly needs to be dragged in the right direction. That gap between expectation and reality, she argues, is exactly where most people go wrong with AI in design.
The fix is not better tools. It is a better approach.
Think Like Tony Stark
To get your donkey working like Jarvis, you need to be more like Tony Stark — methodical, structured, and clear about what you are asking for. Yuliia's framework before touching any AI tool:
Data preparation — identify the key message hiding behind your data
Idea first — decide how you want to show it before you open any tool
Right tool, right task — different AI tools have different strengths
Split the work — be explicit about what you are asking AI to do and what you will handle yourself
AI supports every step, but it cannot own any of them. Give it vague instructions and it will start guessing. It never ends well.
What ChatGPT Is Actually Good At
Rebuilding old charts — screenshot a chart from a presentation where the source file is lost, ask GPT to rebuild it as an editable PowerPoint file. For standard charts, it works cleanly in seconds.
Sketch to slide — photograph a rough sketch or tablet note and ask GPT to turn it into editable shapes and text boxes. Not perfect, but far better than starting from a blank slide.
Colour palettes — ask GPT to generate 12 shades of a specific colour, then use the result directly as a colour reference or grab individual shades with PowerPoint's eyedropper. Faster and more consistent than doing it manually.
Waffle charts and custom grids — describe what you want, reference your data and brand colours, and ask GPT to build an editable shape grid in PowerPoint. Fully adjustable, animatable, and much quicker than building it by hand.
Where AI Consistently Fails — And the Workarounds
Yuliia spent considerable time trying to get ChatGPT to generate a pie chart shaped like a camera lens for a Ukrainian Vogue project. The logic was always wrong. The proportions were off. After many retries, she stopped arguing with it and asked for something different: just generate the lens shape. She then dropped it into PowerPoint, used the real chart as an underlying layer, added fragmentation lines manually, and got exactly what she wanted.
The lesson: if AI will not give you the whole thing, take the best part and finish it yourself. Great visuals do not have to be all or nothing.
For product photography — like needing five cars in consistent style and composition — ChatGPT can get close but will not reproduce real objects exactly and tends to crop images awkwardly. Her fix: take the ChatGPT output to Midjourney, upload the image, resize it to fit the slide, and prompt Midjourney to fill the gaps on a clean background.
Q&A Highlights
How long did the full presentation take to make? About three weeks from start to finish.
What is the hardest data to visualise? It always comes back to the same question: what is the core message behind the data? Without a clear answer to that, no tool — AI or otherwise — will save you.
Final Thoughts: AI Is a Tool. You Are the Driver.
The tools are improving fast. Everything Yuliia demonstrated today may be ten times easier six months from now. But one thing will stay true regardless: AI does not think, it executes. The idea, the message, and the judgement are yours. That is not a limitation — it is the job.
Join the Conversation
What is the trickiest data you have ever had to visualise — and did AI help or make it worse? Share your thoughts below.
Comments