I’ve been testing Notion AI to auto-generate a first draft of product FAQs for our new feature launches. It’s decent for generic structures, but the output feels too vague. It lacks the specific nuance of our actual product capabilities and customer pain points.
I’m feeding it a page with our product specs and a few sample customer support tickets, but the AI still seems to miss the mark. For example, it’ll generate a correct but shallow answer about "integration options" without mentioning our specific pre-built connectors or API rate limits that users always ask about.
Has anyone cracked the code on structuring source material for this? I’m curious about the data-driven approach. Are you creating a dedicated “context database” with linked records of common issues? Or is it better to paste raw, verbatim customer questions as a block before prompting?
I’m trying to optimize for accuracy without having to heavily edit every generated answer. What’s working in your workflows?
✌️
Oh, the struggle of trying to get a generic AI to understand your specific, beautiful snowflake of a product. Been there.
You're on the right track with the support tickets, but you're probably not being ruthless enough in curating them. Throwing a whole page at it is a mistake. The AI gets lost in the noise. What worked for me was creating a separate "FAQ Context" page structured not as raw data, but as explicit Q&A pairs *written by me first*.
Forget the specs document. Instead, take those support tickets and manually write the *perfect* answer you'd want to appear. Do this for, say, your top 20 most common issues. Make each answer brutally specific: name the exact connector, state the exact rate limit, link to the exact doc URL. Then, feed *that* curated list as your primary context, and prompt the AI with "Generate new FAQs in the style and specificity of the examples provided."
It's more upfront work, but the output is 90% there instead of 40%. Otherwise, you're just teaching the AI to be vaguely correct, which is useless.