Anti-patterns
Eight things that kill interview quality.
The eight items below are the recurring quality killers. None of them are exotic. Most are caught by reading the first three real transcripts a blueprint produces - which is itself the first anti-pattern below. Each item names the symptom (what you'll see), the fix, and the chapter that covers it properly.
1. Publishing without running it on yourself
Symptom. The graph has an unreachable branch, the probing depth is too aggressive for the channel, the intro message reads as marketing copy - and you only find out after sending the link to ten real respondents.
Fix. Run the published blueprint on yourself end-to-end. Take the call, type the chat, click the embed. You'll catch four of the next seven items before they ship.
See also. Structuring the interview, Channels, consent, identity.
2. Leading questions
Symptom. Every respondent gives roughly the same answer, in roughly the agent's framing. The themes look strong but the data is the agent's belief, mirrored.
Fix. Rewrite questions to start with "what" or "how" rather than "do you agree" or "would you say". Keep the framing of the respondent's situation in their own register, not the team's.
See also. Designing the agent, Interview techniques - calibrated questions.
3. Stacking two questions in one node
Symptom. The transcript has answers that address one half of the node and ignore the other. The completion_hint can't fire reliably.
Fix. One node, one objective. If a question naturally splits ("when did you last use it, and why did you stop?"), split it into two nodes. Branching, if you need it, sits between them.
See also. Structuring the interview.
4. Techniques that fight each other
Symptom. The agent feels schizophrenic. It mirrors and labels for two turns, then plays devil's advocate, then goes warm again. Respondents disengage.
Fix. Pick a technique stack that fits the interview's posture. Discovery and rapport interviews lean on active_listening, mirroring, labelling, tell_me_more. Stress-testing interviews lean on devils_advocate, socratic, calibrated_questions. Don't mix the two stacks on the same agent.
See also. Designing the agent, Interview techniques.
5. Probing depth mismatched to channel
Symptom. Respondents drop out mid-interview on chat. Or the voice transcript is full of one-line answers because the agent isn't probing.
Fix. deep probing on chat is brutal - the typing burden compounds. Halve max_follow_ups_per_question for chat versus voice. If the interview genuinely needs depth, run it on voice.
See also. Probing depth and follow-ups, Channels, consent, identity.
6. Branching on agent_assessed when structured would be deterministic
Symptom. Two runs with apparently identical answers branch differently. Cohort analysis on the branch is unreliable.
Fix. If the condition is on a typed value (rating, choice, yes/no), capture the value in a form node and branch with a structured condition. Reserve agent_assessed for branches that genuinely require natural-language understanding of an open answer.
See also. Structuring the interview - edges and conditions.
7. No completion_hint on open nodes
Symptom. The agent moves on too early on some runs and probes forever on others. Run length variance is huge for no good reason.
Fix. Write a completion_hint for every open node. It should describe what the agent should have heard, not what to ask. "The respondent has named at least one specific moment when X happened and described what they did in response."
See also. Probing depth and follow-ups.
8. Treating analysis output as ground truth
Symptom. A team decides on a feature based on "Theme 1: users want lower prices, frequency 7/10". They do not read the seven verbatims. The decision is brittle and the team is surprised when the feature lands.
Fix. Themes are summaries, not facts. Read the verbatims for at least the top three themes before acting on any of them. Be willing to discard a high-frequency theme if the verbatims show it's an artefact of the question.
See also. Decision-ready output.
The fastest quality improvement on any blueprint is a single pilot run with yourself as the respondent.
Every senior interviewer running interviews professionally still pilots their guides - they have the reps and they still do it, because the gap between how a question reads on the page and how it lands in a live conversation is the gap a pilot exists to close. Every team that ships a Debriefer interview without running it through themselves first will hit at least three of the items above. The pilot is not a polish step. It is the cheapest, highest-leverage move available to you, and it costs fifteen minutes.