You know the feeling. You took a photo of a beautiful sunset or a mouth-watering spread of BBQ. But when you scroll through your camera roll, it looks flat. It lacks the curated intentionality of the aesthetic feeds you obsess over.
Your camera roll isn’t a failure of memory; it’s a failure of design.
We all want our personal memories to look like polished editorial layouts. So, when AI image tools introduced ‘Skills’—pre-packaged templates that transform ordinary photos into watercolor tickets or crayon-sticker collages—it felt like a magic bullet. But recently, base AI models like GPT, Doubao, and Qwen got terrifyingly good at direct reference-image transfer. You just feed them a photo and a target style, and they spit out a masterpiece.
This raised an uncomfortable question: If the base model can already do style transfer, why bother installing a separate Skill?
I ran a brutal test. I took a boring landscape photo and a cluttered BBQ table shot. I pushed them through four different AI models. Half the time, I used direct reference-image transfer. The other half, I installed a dedicated Skill. The results completely changed how I think about AI creation.
When I used direct style transfer on the landscape, the AI did exactly what I asked—but it hallucinated. It slapped ‘Budapest’ and fake coordinates onto my local park photo. It mimicked the watercolor brushstrokes, but it completely lost the ticket layout. The AI didn’t understand the design system; it just copied the paint.
The dedicated Skill, however, maintained the structure. It kept the photo on top and the watercolor ticket on the bottom. But it still struggled with text, occasionally inventing dates and requiring multiple regenerations to get the typography right.
For the BBQ photo, the exact opposite happened. The direct transfer was fantastic—GPT intelligently removed the messy table background, leaving only the food with a clean crayon texture. The Skill version, however, was a cluttered mess, cramming every single element onto the page without understanding negative space.
Here is the twist nobody is talking about.
We thought installing a Skill was about teaching the AI how to draw. It’s actually about installing someone else’s taste into your workflow.
Most users assume installing a Skill is about gaining a new capability. It isn’t. The base models already have the capability. As these models improve, the bottleneck isn’t generation quality—it’s curation, trust, and aesthetic direction. Skill creators aren’t selling a new way to draw; they are encoding their specific design workflows into a repeatable template.
This is the tension that will define the next year of AI tools. The stronger the base image model becomes, the less a standalone ‘style skill’ matters. Why install a plugin just to get a watercolor effect when you can just upload a reference image? The era of the single-style shortcut is dying.
But in its place, the ‘design recipe’ Skill is being born. What remains valuable are Skills that package an entire designed system—the layout, the motifs, the typography, and the structural logic.
A single style is no longer a product; it’s a freebie. The real value lies in the entire design recipe.
If you just want to rescue a single photo from mediocrity, skip the Skills. Just send a reference image. It’s faster, it’s less friction, and the base models are good enough to nail the vibe. But if you are building a brand, a zine, or a recurring aesthetic where every photo needs to fit a precise, coherent system, you need a Skill.
Just remember to verify the text, dates, and coordinates before you trust the output. The AI might have great taste, but it still lies.
FAQ
Q: If base models can already copy a reference image, why would anyone ever bother installing a Skill?
A: Because copying a style is not the same as applying a design system. A reference image tells the AI 'make it look like watercolor,' but a Skill tells it 'put the photo on top, the ticket on the bottom, use this specific font, and don't invent a location.' It's about repeatability and layout, not just visual generation.
Q: How do I know which method to use for my own photos?
A: If you just want to make a single photo look cool, use a reference image. It's faster and the base models are good enough. If you are building a recurring aesthetic where every photo needs the exact same layout, motifs, and typography, install a Skill.
Q: You said Skills are a distribution play, not a capability play. What does that mean?
A: The AI doesn't need a Skill to know how to draw. The Skill is just a vehicle for a creator to hand you their specific taste and workflow. You're not downloading a new feature; you're subscribing to someone's curated aesthetic.