Meeting Minutes Generator: Saving 2 Hours When Writing Government Minutes
I designed Noms, an AI writing tool that transforms how Singapore government officers write government-style meeting minutes: from a dreaded, hours-long task to a quick edit of dependable first drafts.
1714 hrs
of meeting minutes processed
4.1/5
overall meeting quality rating
112 mins
saved per Notes of Meetings generated
Role
Senior and Sole Product Designer
Owned research, product design, and created quality heuristics that guided our prompt engineering.
Team
2 Engineers, 1 Operations, 1 PM
Timeline
Feb - Jun 2024
Context
Officers spend entire workdays writing government-style meeting minutes
A full workday consumed
Writing a single set of minutes can take 4-6 hours on top of an officer's existing responsibilities.
Precision is non-negotiable
Government minutes are legal records with its own nuances and requirements. They capture decisions, justifications, and who said what: details that inform policy for years.
Prompting was harder than writing
Officers tried ChatGPT and other commercial tools but the effort to prompt "government quality and formatting" was harder than just writing it themselves.
How might we cut hours of documentation work without sacrificing the quality that government minutes require?
How might we cut hours of documentation work without sacrificing the quality that government minutes require?
Research
Turned insights into heuristics for what "good" looks like
I couldn't design for government writing without understanding it. Conducted 12 interviews across agencies, meeting types, and seniority levels and wrote minutes myself to feel the pain.
Prompting
Translated research heuristics into prompt architecture, which underscored the quality of the output.
Mirrored how officers actually write. Officers write minutes in stages, so we built our prompts the same way, with each step targeting specific heuristics.
Continuous iteration with engineers. Generate, compare to real minutes, find gaps, refine. We tested new models, explored prompting techniques, and addressed edge cases together.
Interface Design
Designed for quick verification and effortless correction
Transcript sidebar for instant fact-checking
Officers typically rewind audio recordings to verify who said what. Each generated point links to its transcript source. Click to locate, play audio to verify.
Pre-curated AI edits for quick corrections
Drafts weren't perfect immediately, but officers shouldn't need to craft prompts to fix them. Predefined correction types let officers make quick fixes without leaving the document.
One-click speaker reassignment
Minutes are attributed to specific speakers and become official records. One click reassigns a speaker across the entire document and transcript, no hunting through text.
Joy
A playful brand made a tedious task feel lighter
The name "Noms" comes from Notes of Meetings, but we leaned into the "nom nom" wordplay. Users have emailed us saying the tone was 'refreshing' and 'brought joy to a tedious task.' In a space dominated by sterile productivity tools, a little warmth goes a long way.
Reflections
Define "good" before designing.
AI writing tools require defining exactly what 'good' looks like upfront. The 8 heuristics became standards embedded in both prompts and interface. It became objective criteria for debating changes, not opinions.
Enhance workflows, don't replace them.
Success came from creating writing interactions that suited how officers already worked. Watching officers write minutes gave us rich insights into their process.
Designers should own the prompt layer.
This project taught me to get my hands dirty on the prompt layer. Obsessing over output quality is core UX work now.
What I'd do differently
Earlier user testing of prompt outputs, and exploring collaborative use cases. NOMs often involve multiple officers.
Growing up, I learnt that a warm bowl of noodles and cut fruits is a language of care. So here's some virtual care from me to you, and thank you for stopping by!
