Turning a static content style guide into a reusable AI-assisted workflow for developing clearer, more consistent government services.
While designing and testing digital benefit applications for the Social Security Administration, I noticed a recurring pattern: many usability problems weren't necessarily caused by the interaction design. Applicants could navigate the interface but still struggle to understand the language used in questions, explanations, and instructions.
These were content design problems.
Without access to a dedicated content designer, I began exploring whether generative AI could help me produce clearer, more consistent content that followed SSA's established writing standards.
What started as a relatively simple prompt evolved into a reusable Microsoft Copilot agent that incorporated the agency's content style guide and a set of broader content design principles.
The agent became part of my design workflow, helping me produce content ready for usability testing more quickly while strengthening my own understanding of content design.
The project also changed how I thought about AI-assisted work: the value wasn't simply in generating text, but in creating a repeatable process for developing and evaluating content against established standards.
As our team evaluated benefit applications through usability testing, we encountered problems that couldn't always be resolved by changing the layout or interaction.
Applicants sometimes struggled with the wording of questions, the relevance of requested information, or the explanations accompanying important decisions.
Improving these experiences required more than rewriting sentences. It required an understanding of plain language, information hierarchy, accessibility, and the agency's content standards.
SSA maintained a content style guide that addressed voice and tone, formatting, word choice, and plain-language writing.
However, applying that guidance consistently required time and familiarity with the documentation.
Without direct access to a dedicated content designer, I needed a way to develop and evaluate content while continuing to move quickly through design and usability testing.
My central challenge was to make established content design guidance easier to apply within my everyday UX workflow.
I began investigating whether generative AI could help bridge that gap.
This was an independent initiative that I developed alongside my UX design responsibilities.
I researched prompting approaches, experimented with available generative AI tools, developed and refined instructions, and eventually configured a Microsoft Copilot agent.
I was responsible for determining how the agent should use SSA's content style guide, what general design principles it should follow, and how it should explain its recommendations.
I also used the resulting tool in my own design work, evaluating its suggestions and refining content before usability testing.
Although I shared the prompt and later the agent with colleagues, adoption remained limited. The project's primary demonstrated impact was on my own workflow.
My initial experiment was relatively simple.
I created a prompt based on the voice and tone guidance in SSA's content style guide.
I wanted to explore whether generative AI could help produce language that aligned with the agency's communication standards.
This was a useful starting point, but it represented only a small portion of the guidance available.
The style guide also contained extensive documentation on formatting, word choice, and plain language.
As I continued exploring the approach, I became dissatisfied with how narrowly the initial prompt addressed content design.
Conversations with a content designer at another company also influenced my thinking. They shared information about how they had documented instructions for their own AI agent, introducing me to a more comprehensive approach.
I began reconsidering what the tool needed to do.
Rather than simply producing language with an appropriate voice and tone, I wanted it to apply a broader set of content design standards.
I expanded the prompt to incorporate more of SSA's content guidance.
The goal was to move beyond stylistic consistency and account for additional requirements, including formatting conventions, preferred terminology, and plain-language practices.
As the instructions became more comprehensive, however, the prompt also became increasingly long.
I maintained it in a Word document that I shared with colleagues.
Some of the generative AI tools available to us could accommodate the larger prompt, so the approach was functional.
But it was cumbersome.
Using the prompt required working with a lengthy set of instructions, and the documentation wasn't particularly convenient to distribute or maintain.
The expanded prompt addressed some limitations of the original experiment, but it revealed a different problem.
As Microsoft Copilot became available within SSA, I began exploring how its capabilities could support the workflow.
Technology adoption in government environments can be gradual, so the tools available to me changed over time.
When Copilot agents gained the ability to reference Word documents, I had an opportunity to rethink how the content guidance was organized.
Instead of maintaining everything within a single lengthy prompt, I separated the agent's general instructions from the reference material it needed to consult.
The SSA content style guide became a knowledge source, while the agent's instructions established how it should approach content design tasks.
This created a more reusable structure.
The agent could reference the established style guidance while applying a consistent set of broader content design principles to each request.
I wanted the agent to do more than rewrite content according to a set of stylistic preferences.
Its instructions established expectations for how it should approach content design problems.
These included applying best practices for information architecture, taxonomy, and information hierarchy; advocating for progressive disclosure and plain language; considering accessibility; and developing content suitable for localization.
I also included two requirements that were particularly important to how I wanted to use the tool.
I instructed the agent to explain how its suggestions aligned with its general principles and SSA's content style guide.
Rather than simply accepting generated language, I wanted to understand the reasoning behind a recommendation.
This made the agent more useful as a design aid.
I could examine its suggestions, consider the guidance it was applying, and make informed decisions about whether the proposed content was appropriate.
The explanations also helped reinforce my own understanding of the style guide.
I also instructed the agent not to invent rules when the style guide didn't address a particular issue.
Instead, it needed to acknowledge the gap.
This distinction mattered because the agent was intended to support the application of established standards, not create new organizational policy.
By asking it to identify missing guidance, I could distinguish between recommendations grounded in documented standards and situations requiring additional judgment.
These instructions were intended to make the agent's output easier to evaluate and help me use it as a content design assistant rather than an unquestioned authority.
I used the agent to develop and refine question labels, introductory explanations, and instructional text across benefit application projects.
These were recurring content needs rather than a single isolated writing task.
The agent gave me a consistent starting point for developing language that considered the agency's style guidance and broader content design principles.
Instead of repeatedly interpreting documentation and developing content from scratch, I could use the agent to generate recommendations, review its explanations, and refine the results for usability testing.
The tool didn't replace my responsibility for content decisions.
I still needed to evaluate whether the suggested language fit the applicant's circumstances, the surrounding interaction, and the requirements of the application.
Its value was in supporting that process and helping me move through it more efficiently.
Over time, the agent became a regular part of how I approached content design.
The agent increased my velocity in producing question labels, introductory explanations, and instructional text ready for usability testing.
It helped me move more quickly from identifying a content need to developing language that could be evaluated with applicants.
Although I didn't formally measure the time saved, I experienced a meaningful improvement in my ability to produce content within the design process.
An unexpected benefit was how much the agent helped me learn.
Reviewing its recommendations and explanations made me more familiar with SSA's content style guide and reinforced the principles behind its guidance.
I became better at recognizing recurring content problems and more comfortable making content design decisions without direct access to a dedicated specialist.
The tool wasn't just helping me produce content. It was helping me develop a stronger understanding of the discipline.
I shared the agent with colleagues, but adoption was limited and others used it only occasionally.
While I saw benefits in my own work, I don't have evidence that it produced measurable improvements in content quality or efficiency across the broader organization.
The project demonstrated value within my individual workflow, while also revealing an opportunity to explore broader adoption.
One of the most important lessons was that AI can be valuable as a tool for developing expertise, not just accelerating production.
I initially approached generative AI as a way to help write content when dedicated content design support wasn't available.
Over time, I found that reviewing its recommendations and explanations also strengthened my own understanding of the standards I was trying to apply.
The project also reinforced the importance of structuring AI-assisted workflows around clear expectations.
Providing comprehensive reference material was only part of the solution.
The instructions governing how the agent approached tasks, explained its reasoning, and handled gaps in the documentation were equally important.
I learned that a useful AI assistant needs more than access to information. It needs a clear purpose, an understandable process, and boundaries around what it should claim to know.
Looking back, I would have made a greater effort to share the agent across the organization and establish a process for its continued improvement.
I focused primarily on developing a tool that supported my own content design work rather than treating adoption as a separate design challenge.
Broader use could have helped colleagues produce more consistent, higher-quality content, particularly when dedicated content design support wasn't available.
It also would have created opportunities to improve the agent itself.
Feedback from more designers could have revealed gaps in its guidance, recurring problems with its recommendations, and friction in how people interacted with it.
With that feedback, I could have refined the instructions, made the agent easier to use, and developed a more systematic approach to evaluating its usefulness.
I also would have considered ways to assess its performance more deliberately, rather than relying primarily on my experience using it.
The project reinforced that building a useful tool is only part of creating value. Helping people adopt it, learning from their experiences, and improving it over time are equally important.
This project began with a problem I encountered in usability testing: people were struggling with content, and I didn't have access to the specialized support needed to address those problems as effectively as I wanted.
Rather than treating that constraint as fixed, I explored how emerging technology could help.
The solution evolved incrementally, from a basic prompt to a comprehensive set of instructions and eventually a reusable agent grounded in SSA's content style guide.
Along the way, I developed a workflow that helped me produce test-ready content more quickly and strengthened my own content design capabilities.
The project also changed how I thought about designing with AI.
The most valuable part wasn't the technology itself. It was identifying a meaningful problem, defining how the tool should behave, evaluating its recommendations, and recognizing where further iteration was needed.
For me, the experience demonstrated how experimentation, design principles, and emerging technology could come together to address a practical gap in the design process.
Designing a personalized application experience that established a foundation for modernizing federal benefit applications.
View case studyHelping people navigate consequential retirement decisions through clear content, evidence-based design, and a reusable application framework.
View case study