Context:
With one UX writer supporting all global Prime surfaces, the current model of office hours, limited agent use, and mostly direct involvement cannot scale. Although UX designer partners are empowered to write for Prime CX with resources and guidelines in the Writing for Prime wiki, Prime Design Library, and Prime UX Writing Assistant in Quick Suite, the Prime UX Writing Survey reveals that UX Design scored wiki findability at just 33.33/100, and only 67% of UX Designers actually using the Quick agent. However, the ones who do use the Quick agent give the tool 100/100 satisfaction suggesting AI tools serve to scale UX writing best practices when used.

The results of the UX Writing Survey suggested that embedding AI writing assitance into existing  workflows would provide designers with a tool that meets them where they work, rather than require context switching to standalone tools.

When I drafted a few ideal scenarios for seamless content design, use cases centered around generating copy for blank design elements, either vended from libraries or created by designers; reviewing existing copy in both live Prime CX and what’s been drafted by designers and/or product partners; and applying both net-new copy and recommended improvements directly to designs.

• All scenarios required a tool that exists in the Figma environment—aka, plugins
  
• All scenarios required an LLM to understand and apply UX Writing knowledge to generate and review writing, as well as digest additional context provided by both users and the design environment

I prompted Pippin—an internal chatbot powered by Claude and Amazon Bedrock—to generate the necessary technical documentation for my product vision:
Create the technical archtecture for a Figma plug-in that generates and reviews existing copy for variables (header, subheader, body text, buttons, links) for existing Prime Design Library elements (components, patterns, layouts). The plug-in should be trained on Prime UX writing guidelines and be able to provide alternatives and variants for variable copy, as well as ask clarifying questions when needed to generate the write content for the current context and use case (customer journey stage, target audience demographics, legal requirements)

Kiro was able to quickly create a prototype based on a mock API that could complete the core tasks I intended: generate 3 scored variations for new copy, as well as review and score existing copy along with ranked recommendations for improvement.

• Since it was running off mock API, the copy it created was essentially a collection of phrases hallucinated by Kiro to match the intention of the documentation I provided.
  
• The plugin also struggled to recognize Figma layers, and so it couldn’t always detect placeholder copy, nor could it apply any generated or recommended copy to the actual designs, which significantly limited usability
  

From there, it was a lot of back and forth with Kiro chat to refine the plugin UX for human-in-the-loop interactions. All commands, questions, and clarifications were conducted in natural language—I described what I wanted, didn’t want, or didn’t understand to Kiro, and it took care of the coding. A few key decisions:

• After determining that Kiro didn’t have the authentication to directly read the UXW wiki pages I created, I compiled all writing guidelines into a spreadsheet for Kiro to convert to JSON
• Added a panel for users to add additional context to their generative prompts, including a mix of dropdown menus and checklists with open-ended text input to balance key requirements for messaging with the freedom for users to provide context in their own words
• Enabled the ability to “Request review” from the UX writer in the Review panel, which generates an annotation capturing all the plugin’s suggestions and feedback, and then tags me for review
• Figured out that instructing the plugin to search and read actual Figma layer slot names rather than search for matches to pre-existing components allowed it to detect all existing copy and apply generated copy to designs. This almost means the tool is not limited to existing libraries, and is capable of reading and apply copy to any text layers contained in your Figma file, including newly created designs.
• Modified the UI to make it navigable by adding Back and Clear All buttons; leading the UI with Generate / Review and not showing the context editor unless in Generate mode; and having the plugin retain context even after generating copy, so users can iterate, alter, or repeat context without starting from scratch

Once the tool was more intuitive and effective, I collaborated with the Prime UX Design Technologist to create a cloud app containing my UXW knowledge. This serves as a bridge between the Figma plugin interface and Amazon Bedrock AI, and effectively creates another UX Writing agent that can even be ported into other interfaces our team may create in the future.

After providing me with documentation to set up Toolbox and authenticate in my own terminal, I was also able to access Bedrock and run the connected plugin. This means the plugin is no longer generating new or recommending copy based on stale mock API data—the plugin is now powered by AI! 🚀️

I was then able to provide and iterate on system prompts to train the LLM to “think” and apply writing guidelines like an expert Prime UX writer. I also added a chat field that enables users to further prompt the plugin for refinement, explanations, and iteration on generated copy and reviews.

Generates and reviews copy for both components and newly designed elements right within Figma. This solution serves to 1) effectively eliminate placeholder copy from design development, ensuring any content in mocks comes from Prime UXW-trained resources and that stakeholder reviews remain proactively focused on content strategy rather than reactive to FPO placements, 2) ensure conviction in content design much earlier in development, regardless of the UX writer’s availability or involvement, 3) reduce wait time and number of requests for direct UX writer involvement, and 4) automate enforcement of Prime voice and style, PDL requirements, legal guidelines, and accessibility standards.

Demo: Generate & review for Prime Member Page banner

Demo: Review & chat for Prime Member Page discovery card


Demo: Review & generate iterations for billing risk alerts