I am quite fascinated by the capabilities of the new AI tools that seem to pop up nearly every week. However, I'm also uncertain about how my role might change over time due to AI's impact. That's why I try to find use cases for each new tool to experiment with and evaluate its potential.
The Idea
My goal is to reduce the operational work involved in the typical UX cycle for myself and my team. This includes researching, analyzing, creating, testing, and even writing acceptance criteria.
Where I Started
To begin, I wanted to set up an environment that would make researching and analyzing as easy as possible. I aimed to create several AI agents using well-crafted prompts. My first agent was actually an agent to create agents — sounds like a perfect nightmare loop, but hear me out!
My initial prompt to set up my prompt GPT
This allowed me to easily set up every agent I needed without spending too much time thinking about what to write and what to consider in detail.
My Second Agent: Research Agent
Often, I need an overview of existing insights from scientific research, blog articles, business papers, and more. I wanted GPT to provide links and some keywords describing the content, so I could easily screen it and decide which sources to explore further. I didn't want it to provide me with finished, condensed information initially because there is still room for inaccuracies.
This is how my Research GTP prompt created by my Agent GTP looked like
Parameters and rules
Variety of sources: search for trustworthy, diverse and relevant sources that comprehensively cover the topic.
Source citations: Present the sources in the text in such a way that they are easy to follow and cite.
Topicality: Give preference to up-to-date sources if this is relevant to the topic.
Quality over quantity: Choose high-quality sources over a large number of sources.
Neutrality: Present the information objectively and neutrally, without introducing your own opinions.
Format: Present the information in a structured and clear form.
Tone: Professional, precise and factual.
Additional information:
Topic of the article (to be specified by the user)
Specific aspects or questions to be covered (to be specified by the user)
If available, preferred sources or databases (to be specified by the user)
The final outcome of my first request about improving conversion rates in b2b websites was a list like this:
Implement effective call-to-actions (CTAs):
Key strategies: strategic placement of CTAs (e.g., "above the fold"), use of urgency (e.g., limited time offers), and continuous A/B testing.
Source: NextWinz
Psychological patterns
Use of social proof (social proof)
Key strategies: incorporating customer reviews and testimonials to build trust.
Source: Shopify
Building trust through transparency:
Key strategies: open communication about product information, delivery times and return policies.
Source: VWO
Study on the effect of multistep forms
Result: multistep forms can significantly increase conversion rates by making the process less overwhelming for the user. Example: VentureHarbour increased its conversion rate from 0.96% to 8.1% by implementing a multistep form. Source: VWO
Study on the role of loading times
Result: fast loading times are crucial to avoid abandonment and improve conversion rates.
Source: Sana Commerce
This provided a great head start for diving deeper into specific topics. However, providing functional links sometimes didn't work perfectly, so I occasionally needed to Google the actual links.
My Third Agent: Conclusion Agent
I wanted a concise report of all the information I researched, where I could also add my findings and papers, summarizing a lot of information with actionable items at the end.
My Fourth Agent: UI Creation Agent
The last part I tested was an agent that translates my report into a prompt for an UI creation AI. My idea was that if I have all the insights about my problem and how to address it, I should be able to create initial prototypes for testing, iteratively figuring out the best solution.
For this, I used Vercel. I wanted to be as fast as possible, and Vercel creates variations out of the prompts that I can choose from and iterate upon. However, this process was a bit bumpy.
Outcome of my Insights to UI AI:
Design a product page with the following elements:
Hero Section: Include a clear value proposition and a prominent call-to-action (CTA).
Pricing Tiers: Display pricing tiers with no more than three advantages per column.
Social Proof Section: Showcase testimonials from satisfied customers.
CTAs: Place call-to-action buttons above the fold to encourage quick conversions.
Vercel created multiple designs following each prompt but only as isolated solutions, and it didn’t store them. I saw multiple variations quickly disappear after each task. I realized I had to create them one by one and then merge them at the end. This wasn’t perfect but was still better than tasking someone to create all variations.
The UI is not fancy or sophisticated at all but I was not aiming for the perfect solution. I was aiming for something I can test as fast as possible. And the beauty about it was: It even provided me the code of these pages to build something functional in a very convenient way.
Conclusion
Even though I need to clarify if my workflow works for more complex problems than a B2B selling page, the initial steps already show significant improvements in gathering and condensing information. The UI creation part has potential but is far from perfect. But this is just the beginning. Imagine moving from the discovery phase to the testing phase and back, not in days, but in hours!
By actively using these AI tools and continuously refining their use, we can streamline our workflows and focus more on creativity and strategy I believe. I will for sure test more and will post some experiments I did
Commenti