Prompt Tuning Workbench - Meeting Notes Assistant

Variant A

Baseline prompt

Variant B

Structured prompt

Accuracy 9.2

How well the prompt prevents invented actions or unsupported conclusions.

Action clarity 8.8

How clearly the result separates decisions, owners, and next steps.

Conciseness 8.7

How compact the output remains without dropping important detail.

Before

Typical weak output

"The team discussed roadmap updates, blockers, and next steps. Everyone aligned on priorities and there are some follow-ups. Product will keep working with engineering."

Too vagueNo ownersNo dates

After

Improved output shape

"Decision: ship self-serve import in beta before mobile polish. Action: Maya to deliver API readiness review by Thursday. Open question: should enterprise admins get CSV validation warnings pre-upload?"

Named decisionNamed ownerExplicit open question

Tradeoff notes

What changed

Variant B is longer, but the extra structure sharply improves action extraction.
Strict "do not invent" wording improves trust, but can increase "Unspecified" outputs when transcripts are messy.
Section ordering matters: summary first, actions second, questions last produces better scanability for managers.

Final prompt