The Transcript Problem

extraction as a loop

Part of: Meeting Notes to Action Items

Every meeting ends the same way: someone agreed to do something, and by Monday nobody remembers who, or by when. You have a wall of transcript text and you need a short, honest list of who owes what. That gap is the whole product. What we're building By lesson 8 you'll have a tool that takes a raw transcript and returns action items, each with an owner , a task , and a due date , grouped by person. It's the same six-step loop every AI product follows: get the transcript in, wrap it in a prompt, call the model, parse the reply into structured items, verify them, and print the report. The special step here is parse . A summarizer turns text into shorter text. We turn text into structured data : a list of records your code can sort, group, and count. That's the skill this whole track calls extraction. Why not just eyeball it A ten-minute meeting is a thousand words of noise: greetings, tangents, half-decisions, jokes. The action items are three or four sentences buried in it. A human skimming misses some and mis-remembers owners. The model reads the whole thing every time and pulls the commitments out consistently, which is exactly the boring, repeatable job it's good at. The real cal

Challenge: Count the Commitments