Problem
Writing Notes of Meetings (NOMs) is a common pain across the whole of government. It's tedious, time-consuming, and highly essential. It takes hours of an officer's time to ensure NOMs are accurately captured whilst adhering to government requirements (style, tone, formatting), all on top of their regular day-to-day tasks. Officers couldn't use conversational AI tools to automate this task because prompting a "government-quality" document was harder than writing it themselves.
My Role
As the sole designer (with 2 Engineers, 1 Product Operations, 1 PM), I owned the end-to-end design process. Besides conducting research and designing the product, I developed quality heuristics that became essential for collaborating with engineers on prompt engineering and evaluation.
Research: Understanding Government Meeting Minutes
I spoke to 12 participants who wrote minutes for different agencies and meetings of different formalities and even wrote minutes myself to understand how to write good government minutes. Key findings:
- 📇 Government minutes are historical records: must capture decisions and alternatives with justification
- ☑️ Distilled 8 heuristics of good minutes: Content Identification, Conciseness, Speaker Attribution, Flow, Structure, Language/Tone, Accuracy, and Formatting.
Prompt Engineering: From Research to Prompts
Officers quickly rejected the tool if initial output quality was poor, so quality was essential. I worked closely with engineers to translate our quality heuristics into the prompt architecture, splitting generation into smaller steps that each perfected different heuristics, mimicking an officer's thinking. Through iterative refinement, I compared our generations against real minutes, identified areas for improvement, and tweaked repeatedly while adapting to model changes and new prompting techniques.
Interface Design: Creating a Familiar Writing Interaction
1. Quick ability to reference transcripts
Officers often had to refer back to the original meeting recording or transcript for fact-checking. When users write NOMs, they typically have audio recordings or raw notes beside them, rewinding to check details when confused about who said what. I built a transcript beside the NOMs where you could click to see where each point corresponds to and play the audio for quick fact-checking, complementing their existing workflow.
2. Edit with AI
Generated first drafts weren't perfect immediately. I empowered users with AI editing functionality that helped with quick pre-curated edits.
3. Editing speakers
Since meeting minutes are directly attributed to the speaker, it's important to be able to assign the correct speaker. In one click, you can assign the correct speaker across the entire document and transcript.
4. (Bonus) Sprinkling Delight
Since writing NOMs was already perceived as such a chore, we decided to make the process more enjoyable for users with a playful approach. The name "Noms" was derived from Notes of Meetings but led to fun wordplay with "nom nom" (eating). Users have emailed us saying the playful tone was refreshing and brought joy to an otherwise tedious task.
1714 hrs
of meeting minutes processed4.1/5
overall meeting quality rating112 mins
saved per Notes of Meetings generatedKey Learnings
Designing AI writing tools requires defining exactly what "good" looks like and embedding those standards into both prompts and interface. Success came from guaranteeing dependable first drafts through quality-focused prompt engineering, then creating writing interactions that suited how officers were already working. As a designer, this experience taught me to get my hands dirty on the prompt layer as obsessing over output quality is also core UX work now.
