4 ms·
I'm curious about the guardrails here. In my experience trying to use LLMs for user research, they tend to be "yes man" often hallucinating features or agreeing
by lovrok23 10mo ago
I'm curious about the guardrails here. In my experience trying to use LLMs for user research, they tend to be "yes man" often hallucinating features or agreeing to user requests that aren't actually on the roadmap just to keep the conversation flowing.
how do you constrain the agent to stick strictly to the facts of the product hypothesis without making stuff up to please the potential customer?
- Matzalar 10mo agoWe ran into the same issue early on. Our fix was to lock each agent to a small JSON snapshot of the idea (no other knowledge), plus strict response templates. They can only ask questions, never describe features or promise anything. If a user asks for something outside scope, the agent replies with “not in the current hypothesis, why is that important to you?” rather than making stuff up. We also have a human review step before anything goes live.
- KurSix 10mo agoIf you have human review for every action, then this isn't scalable software, it's consulting. Either you'll eventually remove this step for scale (and get banned by platforms), or your product will have to be very expensive (to pay for the reviewers' time). Maybe it's worth keeping AI only for the analysis part and leaving the communication to humans?