This agency is building Custom Widgets with a little help from AI

July 11, 2025
0 minute read

As an agency owner, Callum Wells is all about pushing creative boundaries. He started building websites on the side about five years ago, before taking the leap from marketing professional to full time agency owner at the beginning of 2024. Since then, he’s raced towards the limits of nearly every website builder he’s touched.


Callum deeply values thoughtful, innovative design. Frequently inspired by unique website elements like interactive elements and engaging content displays, he often finds himself asking, “Could I build something similar on Duda?” It’s precisely this curiosity that has led him to incorporate artificial intelligence into his workflow - particularly to solve the “final mile” challenge in web design.


No website builder can realistically support every conceivable design combination. There’s always a trade-off between customization and usability, too many options would overwhelm users. While many website builders aim to get designers “close enough,” his ambition is to bridge the gap completely.


An AI-powered widget factory


These days, Callum wears a few hats. He’s an agency owner, a marketing consultant, and a developer. Although, developer may not be the correct terminology. Rather, the way he toes the line between writing code and instructing AI is more akin to a conductor, orchestrating various artificial intelligence tools to create something great.


You see, it’s not enough to open up ChatGPT and ask for a widget. He frequently emphasizes the human skill behind AI tools.


"I like to compare the use of AI to chefs," he said. "Some chefs hold Michelin stars, others don’t. Generally, Michelin-star chefs produce superior dishes. Similarly, AI models produce the best results when used skillfully by someone who understands their nuances. Some people are better than others, but it is a learnt skill."


His agency, WebHero, actively uses multiple AI models, each chosen for their unique strengths. Claude, particularly following their upgrade to Claude 4 Opus, stands out to him as one of the market's top models for generating code. Alternatively, he prefers ChatGPT for tasks like copywriting, with numerous other models playing key roles as well in various aspects of his workflow.


With the right tool, and the right guardrails, all that’s left to do is prompt. This part, too, is not as easy as it sounds. Callum found the process of prompting new widgets into existence was one that involved a lot of trial and error, often requiring multiple iterations. Models cannot intuit what you mean—as lovely as that might be. Context is everything.

A great looking widget that functions well on desktop, for example, may not be responsive on mobile if you don’t specifically prompt for that functionality.


"Being explicit is key," Callum warns. He finds it helpful to explain to the model exactly how the code should function on Duda. In addition to his thorough prompting, he has Google Docs files filled with support and API documentation to provide the models with even more context.


Developing this context has been a labor of love. Over the past twelve months, Callum has invested hundreds of hours perfecting these prompts and refining his interactions with different models. He's found that explicitly instructions and detailed context significantly enhance the output of these models.



Top Tip
Did you know? Agencies can generate Custom Widgets directly within Duda using the AI Assistant, with all of that important platform context baked right in.


All of this context comes at a price, which he pays in the form of tokens. Adding platform documentation, dozens of guardrail prompts, and explicit descriptions per widget can chew through usage allotments fast—particularly on free plans, and especially when iterating through designs.


Of course, Callum is not using the free version of these tools. He invests significantly in his process, to the tune of nearly $600 per month. Some of those tools include:


  • Claude Max
  • Manus Pro
  • Super Grok
  • Perplexity Pro
  • ChatGPT


These tools are his orchestra.


Widgets, widgets, widgets


Callum likes to, in his own words, "think beyond what is possible," when it comes to website design. When Callum encounters an engaging or unique website element, his first instinct is always the same: take a screenshot, pop it into Claude, and quickly generate exactly what he envisioned using carefully refined prompts. Within seconds, AI produces functional, responsive components that traditionally required extensive custom coding.


Through this efficient workflow, Callum has accumulated a diverse collection of custom AI-generated components, including infinite scrollers, swipeable containers, draggable stickers, expandable cards, and scaling hamburger menus—to name a few.


Callum shared the generated code with us below.


Expandable cards



Callum's "Expanding Cards" widget adds click-to-expand card animations to your site. You can get the full code, alongside his write-up, at AgencyGenius.co.


Scaling hamburger menu




Callum's "Scaling hamburger menu" widget adds an elegant, animated menu to your site. You can get the full code, alongside his write-up, at AgencyGenius.co.


What does he do with all of these widgets?


The best part of working with AI is how conversational it is. You can take the embedded code and paste it into another AI model (Callum recommends Claude), then customize it to your liking through natural language prompts.

Again, the key is to provide clear, contextual instructions specific to each component. For example, with his Expandable cards widget, you are likely going to need some kind of image or video for each card. He recommends uploading these assets to Duda, copying the URL, then sharing it with Claude to modify his provided code.

Of course, a big risk of working with any AI model is hallucinations. He recommends concluding your prompt with some kind of guardrail, such as:

"Make no other changes beyond what’s specified, and ensure the output remains neatly contained within an HTML <div> class, with all CSS and JavaScript included inline."


Top Tip
If you want to place multiple custom components on the same page, ask your AI assistant to change the name for the second iteration. This will prevent styling conflicts.


Once you start to get a grasp of the finesse required to generate high-quality website elements, you might find the process a little addictive. So addictive, in fact, that before long you could become flush with widget after widget. At least, Callum did.

That's why he chose to share his work.
AgencyGenius, Callum's repository for custom website elements, was born from Callum's direct experience and desire to empower others. Members gain access to a curated vault of pre-built, AI-ready components. They simply copy these components and prompt their own AI models with natural language to rapidly adapt and customize designs.


His work comes in the form of either copy-and-paste HTML, or more native Custom Widgets fully integrated into the Duda platform.

An example of how AgencyGenius distributes code to its members.


The finishing touch on AgencyGenius is his educational component, which includes clear demonstrations exploring how to build components from scratch, as well as a robust prompt library.


Callum describes AgencyGenius as a toolbox he would’ve liked to have had five years ago when he first started building websites on the side. In addition to website elements, he offers development resources, tools, inspiration, and a slack community to his members—all intended to help them grow.




The future of AI-powered agencies


More and more, agencies are exploring fulfilment and productization. TheCamel, a longtime Duda Expert, offers a similar, albeit more comprehensive, toolbox with their "All Access" program. CommonNinja, as well, offers many widgets, both free and paid, on their website as well as on Duda’s app store.


What makes AgencyGenius so remarkable is that the entire initiative is truly a one-man show. At Duda we talk a lot about the efficiency gains to be had with AI. We write about the power of AI, host webinars that discuss it, and develop new features based on it. Yet, Callum truly embodies it. He is living proof of what an “AI Agency” can be.


"I'm far too humble to say that I have truly done this on my own," he said. "The fact is that this wouldn't have been possible without a platform like Duda and their new Editor 2.0 is an incredible solution that, in my opinion, all agencies should be using. Not just because of its White Label benefits, but because it is truly the most reliable platform of all the ones I've used. WebHero is beginning the transition of all websites to Duda. I'm incredibly proud that AgencyGenius is 100% built on Duda… with a little bit of help from AI."


Two years ago, Duda’s CEO, Itai Sadan, penned an article for this blog titled “Beyond the hype: What AI means for agencies.” In that piece, Itai stated “Fortunately, I don’t believe that AI will replace agencies. While AI is indeed reshaping how SMBs run their businesses, the value of the agency is here to stay. However, agencies will need to restrategize the way they support their clients and create value.” During our interview, Callum made a similar statement, claiming that “AI won’t replace agencies, [but it will] push [them] forward. The most successful businesses will be leveraging AI in every aspect of their [workflows.]”


We’re finally starting to see what that next generation of agency might look like.


Ready to start building your own widgets?


Callum’s advice for agencies exploring widget creation with AI is straightforward: Stay informed. AI is evolving rapidly, with new models and approaches appearing frequently. Keep up with podcasts, newsletters, and industry thought leaders - some of Callum’s personal favorites being:



Most importantly, he encourages designers and agency owners to approach AI primarily as a creative amplifier. Focus first on design possibilities rather than purely functional concerns, and you’ll consistently produce better, more compelling results.


Have you experimented with AI-generated widgets? Share your experiences with the Duda Community on Facebook!


Headshot of Shawn Davis

Content Writer, Duda

Denver-based writer with a passion for creating engaging, informative content. Loves running, cycling, coffee, and the New York Times' minigames.


Did you find this article interesting?


Thanks for the feedback!
By Shawn Davis April 16, 2026
Website builder analysed 69M AI crawler visits across over 850,000 websites in February 2026 to determine key trends and characteristics that increase local AEO
By Shawn Davis April 1, 2026
Core Web Vitals aren't new, Google introduced them in 2020 and made them a ranking factor in 2021. But the questions keep coming, because the metrics keep changing and the stakes keep rising. Reddit's SEO communities were still debating their impact as recently as January 2026, and for good reason: most agencies still don't have a clear, repeatable way to measure, diagnose, and fix them for clients. This guide cuts through the noise. Here's what Core Web Vitals actually measure, what good scores look like today, and how to improve them—without needing a dedicated performance engineer on every project. What Core Web Vitals measure Google evaluates three user experience signals to determine whether a page feels fast, stable, and responsive: Largest Contentful Paint (LCP) measures how long it takes for the biggest visible element on a page — usually a hero image or headline — to load. Google considers anything under 2.5 seconds good. Above 4 seconds is poor. Interaction to Next Paint (INP) replaced First Input Delay (FID) in March 2024. Where FID measures the delay before a user's first click is registered, INP tracks the full responsiveness of every interaction across the page session. A good INP score is under 200 milliseconds. Cumulative Layout Shift (CLS) measures visual stability — how much page elements unexpectedly move while content loads. A score below 0.1 is good. Higher scores signal that images, ads, or embeds are pushing content around after load, which frustrates users and tanks conversions. These three metrics are a subset of Google's broader Page Experience signals, which also include HTTPS, safe browsing, and mobile usability. Core Web Vitals are the ones you can most directly control and improve. Why your clients' scores may still be poor Core Web Vitals scores vary dramatically by platform, hosting, and how a site was built. Some of the most common culprits agencies encounter: Heavy above-the-fold content . A homepage with an autoplay video, a full-width image slider, and a chat widget loading simultaneously will fail LCP every time. The browser has to resolve all of those resources before it can paint the largest element. Unstable image dimensions . When an image loads without defined width and height attributes, the browser doesn't reserve space for it. It renders the surrounding text, then jumps it down when the image appears. That jump is CLS. Third-party scripts blocking the main thread . Analytics pixels, ad tags, and live chat tools run on the browser's main thread. When they stack up, every click and tap has to wait in line — driving INP scores up. A single slow third-party script can push an otherwise clean site into "needs improvement" territory. Too many web fonts . Each font family and weight is a separate network request. A page loading four font files before rendering any text will fail LCP, especially on mobile connections. Unoptimized images . JPEGs and PNGs served at full resolution, without compression or modern formats like WebP or AVIF, add unnecessary weight to every page load. How to measure them accurately There are two types of Core Web Vitals data you should be looking at for every client: Lab data comes from tools like Google PageSpeed Insights, Lighthouse, and WebPageTest. It simulates page loads in controlled conditions. Lab data is useful for diagnosing specific issues and testing fixes before you deploy them. Field data (also called Real User Monitoring, or RUM) comes from actual users visiting the site. Google collects this through the Chrome User Experience Report (CrUX) and surfaces it in Search Console and PageSpeed Insights. Field data is what Google actually uses as a ranking signal — and it often looks worse than lab data because it reflects real-world device and connection variability. If your client's site has enough traffic, you'll see field data in Search Console under Core Web Vitals. This is your baseline. Lab data helps you understand why the scores are what they are. For clients with low traffic who don't have enough field data to appear in CrUX, you'll be working primarily with lab scores. Set that expectation early so clients understand that improvements may not immediately show up in Search Console. Practical fixes that move the needle Fix LCP: get the hero image loading first The single most effective LCP improvement is adding fetchpriority="high" to the hero image tag. This tells the browser to prioritize that resource over everything else. If you're using a background CSS image for the hero, switch it to anelement — background images aren't discoverable by the browser's preload scanner. Also check whether your hosting serves images through a CDN with caching. Edge delivery dramatically reduces the time-to-first-byte, which feeds directly into LCP. Fix CLS: define dimensions for every media element Every image, video, and ad slot on the page needs explicit width and height attributes in the HTML. If you're using responsive CSS, you can still define the aspect ratio with aspect-ratio in CSS while leaving the actual size fluid. The key is giving the browser enough information to reserve space before the asset loads. Avoid inserting content above existing content after page load. This is common with cookie banners, sticky headers that change height, and dynamically loaded ad units. If you need to show these, anchor them to fixed positions so they don't push content around. Fix INP: reduce what's competing for the main thread Audit third-party scripts and defer or remove anything that isn't essential. Tools like WebPageTest's waterfall view or Chrome DevTools Performance panel show you exactly which scripts are blocking the main thread and for how long. Load chat widgets, analytics, and ad tags asynchronously and after the page's critical path has resolved. For most clients, moving non-essential scripts to load after the DOMContentLoaded event is a meaningful INP improvement with no visible impact on the user experience. For websites with heavy JavaScript — particularly those built on frameworks with large client-side bundles — consider breaking up long tasks into smaller chunks using the browser's Scheduler API or simply splitting components so the main thread isn't locked for more than 50 milliseconds at a stretch. What platforms handle automatically One of the practical advantages of building on a platform optimized for performance is that many of these fixes are applied by default. Duda, for example, automatically serves WebP images, lazy loads below-the-fold content, minifies CSS, and uses efficient cache policies for static assets. As of May 2025, 82% of sites built on Duda pass all three Core Web Vitals metrics — the highest recorded pass rate among major website platforms. That baseline matters when you're managing dozens or hundreds of client sites. It means you're starting each project close to or at a passing score, rather than diagnosing and patching a broken foundation. How much do Core Web Vitals actually affect rankings? Honestly, they're a tiebreaker — not a primary signal. Google has been clear that content quality and relevance still dominate ranking decisions. A well-optimized site with thin, irrelevant content won't outrank a content-rich competitor just because its CLS is 0.05. What Core Web Vitals do affect is the user experience that supports those rankings. Pages with poor LCP scores have measurably higher bounce rates. Sites with high CLS lose users mid-session. Those behavioral signals — time on page, return visits, conversions — are things search engines can observe and incorporate. The practical argument for fixing Core Web Vitals isn't just "because Google said so." It's that faster, more stable pages convert better. Every second of LCP improvement can reduce bounce rates by 15–20% depending on the industry and device mix. For client sites that monetize through leads or eCommerce, that's a revenue argument, not just an SEO argument. A repeatable process for agencies Audit every new site before launch. Run PageSpeed Insights and record LCP, INP, and CLS scores for both mobile and desktop. Flag anything in the "needs improvement" or "poor" range before the client sees the live site. Check Search Console monthly for existing clients. The Core Web Vitals report surfaces issues as they appear in field data. Catching a regression early — before it compounds — is significantly easier than explaining a traffic drop after the fact. Document what you've improved. Clients rarely see Core Web Vitals scores on their own. A monthly one-page performance summary showing before/after scores builds credibility and makes your technical work visible. Prioritize mobile. Google uses mobile-first indexing, and field data shows that mobile CWV scores are almost always worse than desktop. If you only have time to optimize one version, do mobile first. Core Web Vitals aren't a one-time fix. Platforms change, new scripts get added, campaigns bring in new widgets. Build the audit into your workflow and treat it like any other ongoing deliverable, and you'll stay ahead of the issues before they affect your clients' rankings. Duda's platform is built with Core Web Vitals performance in mind. Explore how it handles image optimization, script management, and site speed automatically — so your team spends less time debugging and more time building.
By Ilana Brudo March 31, 2026
Vertical SaaS must transition from tools to an AI-powered Vertical Operating System (vOS). Learn to leverage context, end tech sprawl, and maximize retention.
Show More

Latest posts