Writing the answers an AI assistant reads
Nearly every bad answer traces back to the content behind it, not the model in front of it.
· 1 min read
One topic per card
A single page covering shipping, returns and payment gets retrieved for all three and answers none of them well. Split it until each piece answers one question.
Numbers, not adjectives
"Fast delivery" is unusable. "2–3 working days, ₪25, free over ₪200" is an answer. Every threshold, price and day count you leave vague becomes a place the assistant either refuses or guesses.
Write the edge case down
The unwritten exceptions — the customer who is always given free shipping, the item that cannot be returned — are exactly what the assistant will get wrong in public.
Never write "answer, then hand off"
A rule that says both usually produces only the handoff, and the customer loses the answer. Decide which one the situation deserves.
Keep one language per answer
Mixed-language content produces mixed-language replies. Write each answer in the language the customer will read it in.
Write for the question, not the page
Website copy is written to be browsed; assistant knowledge is written to be retrieved. Give each piece a title that matches how a customer would ask — "how long does delivery take" beats "logistics" — because the title is usually what decides whether it is found at all.
Fix from transcripts, not from imagination
The fastest way to improve an assistant is to read what it got wrong yesterday and write the missing fact. Every hour spent guessing at content is an hour not spent on the questions customers demonstrably asked.