Skip to content
JD Teach AI
Back to blog
Coaching & MethodJean-Dominique Casanova

Meeting notes to actions: a verifiable AI exercise

A fictional fixture, instruction, reference output and six tests for checking a workflow.

Meeting notes to actions: a verifiable AI exercise

This teaching case follows three fictional notes, one instruction and a reproducible check. It shows how to prepare a workflow for human review. The reference output and deliberate mistakes were written for this exercise; no client outcome or time saving is claimed.

Download the complete worksheet · Download the JavaScript prototype

Input: three notes, three situations

SourceFictional note
N1Camille prepares the mockup before September 25, 2026.
N2Alex checks the links. The date is to be agreed.
N3The publication planned for September 24, 2026 is cancelled.

The workflow is notes → proposed extraction → human review → approved table. This exercise ends at the table. It sends no messages and creates no tasks in another tool.

An instruction to try

Using only notes N1 to N3, extract active actions in a table with action, owner, deadline and source. Do not invent an owner or date. Write “to confirm” when a date is missing. Exclude cancelled actions. Keep the source identifier and sentence. End with details to confirm. Do not trigger any external action.

Use the assistant of your choice. Record the tool, displayed model, date and instruction so you can compare your runs.

Reference output and review

This reference was written manually from the fixture above.

SourceActionOwnerDeadline
N1Prepare the mockupCamilleSeptember 25, 2026
N2Check the linksAlexTo confirm

Keep the source sentence alongside each row. N3 is excluded because the action is cancelled. The link-checking deadline still needs to be confirmed with the person responsible.

Check five things: exactly two active actions, correct owners, the exact N1 date, no invented N2 date and faithful sources. A missing action, cancelled action or invented date blocks acceptance.

Unblock the prototype: expose a failure

The standalone JavaScript file contains a deliberately incorrect candidate: an invented date for N2 and an added N3 action despite its cancellation. This is not an observed model output. The validator compares normalized fields with the approved reference and reports:

N2: incorrect deadline
Unexpected action: N3

The correction removes N3 and leaves N2’s deadline empty. The checks also cover a missing action, a duplicate, a changed deadline and malformed output.

With Node.js 20 or later, save the file and run:

node --test meeting-actions-demo.mjs

The six tests check this closed fixture, not the meaning of arbitrary notes. Change the reference deadline without changing the output to see a failing comparison. Then update the output and rerun. This is a concrete starting point for coaching on vibe coding.

Measure your run and hand it over

First extract the actions manually, then use the instruction on the same notes. The worksheet includes a blank log for extraction time, correction time, errors and human approval. Add extraction and correction time before comparing duration. No timed results are published here.

A small synthetic fixture cannot establish business ROI. For a real process, define a period, case count, consistent quality criteria and review costs. The workflow selection guide helps frame that decision.

For handover, give another person the worksheet, change a deadline and ask them to explain which check fails. Keep the instruction, fixture, reference and log together.

Choose the next step

Use individual coaching to learn through your own blocker. Use the team workshop to practise a shared process together. For a delegated implementation with agreed acceptance criteria, describe your process in the project brief.

Share

Read next