Developer improves AI scrum assistant accuracy through structured prompts and team input

A developer built an AI assistant using OpenAI's API to extract action items and blockers from unstructured daily scrum notes. The initial implementation with GPT-3.5 achieved only about 60% accuracy, frequently missing items or hallucinating nonexistent ones. Improvements included detailed prompting instructions, providing example outputs, and upgrading to GPT-4o, which increased accuracy but didn't solve all reliability issues. The final solution combined encouraging team members to add minimal structure to their notes with implementing robust validation for the AI's output.