Interview Agent — adaptive questioning quality

What users said

• They said the current flow does not adapt enough once the first answer comes in.

• They explicitly wanted the tool to generate follow-up questions while the interview is happening.

• They gave the example that if a user says “no,” the agent should not stop there, but ask what tools they prefer or what alternative revenue streams or options matter, depending on interview context.

• They also liked that a human interviewer could manually toggle or intervene if they disagreed.

Underlying requirement

• The Interview Agent needs to behave less like a fixed script and more like a context-aware interviewer that can inspect the meaning of the previous answer, identify missing information, and ask the most useful next question.

Implementation guidance for LLM agents

• Add a follow-up generation layer after each answer:

  • classify the response type: positive / negative / ambiguous / incomplete / surprising

  • identify what key information is still missing

  • generate 1 best next question, not 3–5 options

• Add branching logic for short answers:

  • if answer is “no,” ask why, what alternative they use, what would make them change, and what consequence that has

  • if answer is vague, ask for example / frequency / impact

  • if answer reveals a new issue, pivot into that issue before returning to the script

• Preserve interview goal:

  • follow-ups should still serve the campaign objective, not wander

• Add interviewer control:

  • accept / edit / replace generated follow-up

  • allow manual override without breaking interview state

• Add prompt-level instructions for depth:

  • “Do not accept a one-word answer when a richer answer is needed to make the output actionable.”

• Add memory/state:

  • store answered topics, unresolved threads, contradictions, and named issues

  • use this to avoid repetitive or shallow follow-ups

Acceptance criteria

• After a shallow answer, the system asks a relevant probing question.

• After a negative answer, the system asks for alternative preference or root cause.

• The agent can switch from the planned script to an emergent thread and later return.

• Human operator can override the next question at any step.