A One-Day Spike for the AI Assistant
Work from the brief alone. Write what you would produce — actors, requirements, data, a first slice, your unknowns — before opening anything below it; the exercise stops working if you read the reality first.
A teammate proposes a two-week prototype of an AI assistant over company documentation "to see if it works". Turn it into a one-day spike: the question, the prediction, the smallest setup that answers it, and what the result would and would not tell you.
Your attempt
Nothing here is checked. It exists so that the reveal below is a comparison rather than a reading.
Building the two-week version with a chat UI so stakeholders can try it. It demos well, the stakeholders love it, it is deployed the following week, and nobody ever measured whether retrieval finds the right page — so the assistant confidently answers from the wrong document, in production, with a UI that looks finished.
Read this even if you are confident. It is here rather than behind a button because it is the move most people actually make, it produces things that look like a project, and its cost arrives when the first hard requirement has nowhere to go.