Can Skills Make a Small AI Model Perform Like a Frontier Model?
November 10, 2026 · 3 min read
We often attribute the difference to model size, training data, reasoning capabilities, and compute. And rightfully so. But I’ve been thinking about another possibility…
What if some of the things we consider model intelligence are actually missing instructions, expertise, and guidance? What if a smaller model could perform considerably better at specialized tasks simply by giving it access to the right knowledge and procedures?
This led me to an experiment aimed at achieving Claude-model-level UI;
The Hypothesis:
When we ask an AI model to perform a specialized task, we often expect it to figure out everything independently. For example, asking a model to build a beautiful SaaS landing page requires more than frontend development knowledge. A model might know how to write React or CSS, but that doesn’t necessarily mean it knows how to produce exceptional design. Without sufficient guidance, it has to rely heavily on its existing knowledge and inference. This is where AI Skills become interesting.
Skills are essentially packages of specialized instructions, procedures, domain knowledge, and workflows that an AI agent can retrieve when needed, they don’t retrain the model or change its underlying parameters. My hypothesis was that providing these skills could significantly improve the output of a smaller model.
The experiment:
For this experiment, I chose Muse Spark 1.3-contributor because I wanted to test something practical rather than use an artificial reasoning benchmark. So I decided to make it build a “premium landing page for a fictional AI SaaS product”. I gave it two attempts:
Attempt 1: Without specialized skills:
I launched Muse and gave it a detailed prompt, allowing it to research the web for inspiration and build the best landing page it could. I wanted to see what it could accomplish using its existing capabilities. The prompt:
Build a premium, production-quality landing page for a modern AI SaaS product.
You have creative freedom over the design direction, layout, typography, colors, animations and overall visual identity.
Before implementation:
- Research the web for inspiration from exceptional SaaS landing pages.
- Study modern design trends and identify effective UI patterns.
- Look at products like Linear, Vercel, Raycast, Stripe and other well-designed SaaS products.
Requirements:
- Make it visually exceptional.
- Avoid generic AI-generated website aesthetics.
- Establish a strong visual identity.
- Use thoughtful typography, spacing, hierarchy and composition.
- Make the page responsive.
- Include realistic content and functional interactions.
- Pay attention to every section, not just the hero.
Don't settle for a basic template. Make the strongest landing page you can produce with your current capabilities.
Implement the actual working website.
result:

Attempt 2: With specialized skills (https://www.ui-skills.com):
For the second attempt, I introduced the UI Skills CLI, a collection of specialized design-engineering skills. Installation steps i used:
- ) Installed skills using bun package manager(npx works too)
bunx ui-skills
2.) I created a directory for the skill inside Muse’s global skills directory:
mkdir -p ~/.agents/skills/ui-skills-root && \
bunx ui-skills start > ~/.agents/skills/ui-skills-root/SKILL.md
3.) I then verified that Muse could discover the skill:
muse skills list
4.) After launching Muse code, I invoked the skill:
/ui-skills-root
The model successfully loaded the skill. Then i gave it thesame prompt as above and the result:


To me, Both look slop, but very close to what claude opus 5.5 gives you by default. Which is a significant improvement from the first result.
What does this mean?
I don’t think this experiment proves that skills can turn a smaller model into a frontier model. That would be an irresponsible conclusion from just two generations of one landing page. However, it does demonstrate something worth investigating; The quality of an AI agent isn’t determined by its underlying model alone. Its surrounding ecosystem matters too. Another way of improving ai generated ui is giving it some reference or inspiration to folow.
This is my first article here, and thanks for reading, have a wonderful one 💌