My work 

The context

Our marketing team was experimenting with AI tools to speed up content work. I wanted to see if a custom Gemini Gem could help anyone at Wellhub write better, on-brand copy without needing a UX writer in the room.
 

My role

I led the Gem from concept to internal prototype. I defined the use case, wrote the system instructions, ran the testing, and coordinated with brand leadership for the internal launch.
 

Tools used

Gemini, Claude, Google Docs
 

The problem

If an AI agent can essentially do anything, defining a singular purpose was the first obstacle. Should it build full web pages? Run a message house intake? Act as a glorified editor? And once the use case was pinned down, harder questions came up. What's the ideal output format? How should the AI sound and behave? How do we keep outputs consistent and accurate? What shouldn't it do?
 

The research

I used rapid iterative testing to work through those questions. Testing revealed a key insight: Gemini Gems aren't ready for a complex set of tasks. A single task with a clear format is when you start to see the magic.
 

The solution

Think big. Start small. I set the Gem on a path to be an on-brand copy helper, offering on-demand assistance to anyone writing for Wellhub. From there, I aligned with senior leaders to set a timeline for an internal launch and bring in more participants from our brand team for testing.
 

Key contributions

• Defined the agent's purpose and the boundaries of what it should and shouldn't do

• Wrote and iterated on all system messages that shaped behavior, tone, and decision-making

• Sourced reference documents and benchmarks to ground outputs in trusted sources

• Defined performance criteria to evaluate quality, consistency, and alignment with intended use

• Documented all tests and tracked results to guide ongoing improvements

• Synced with our Senior Brand and Creative Director to align the Gem with our voice and standards

• Presented the prototype to the full brand team

• Maintained a live doc to keep teams aligned and progress visible
   

The result

Based on evaluation by 8 participants, the Gem was an exciting step forward but not yet ready to share broadly across the company. 6 out of 8 participants rated it "Good." Most brand team participants said it would be great for inspiration, while non-writers said they felt comfortable using the outputs as is.
 

The impact

Open questions that came out of testing: Is the content customer-facing ready? Is the Gem foolproof from hallucinations? Are the rules in the system instructions too restrictive? These questions will shape how we approach the next iteration.
 

What's next

Continue fine-tuning based on the latest feedback. Duplicate the Gem on other models or GPTs to evaluate which platforms perform better. Survey marketing and product orgs to learn what types of agents would be most useful, so we can offer a collection that serves a variety of tasks and use cases.
 

The takeaway

Gems can be a powerful tool when given the right data, instructions, and development. But building an AI agent like this takes significant effort to govern and keep outputs consistently high-quality, accurate, and on-brand.