Interview techniques

A technique in Debriefer is a behaviour the agent layers into its probing. You configure techniques as an array on the agent - see the create_agent reference - and the agent draws from that list when deciding how to respond to whatever the respondent just said. There are twelve enum values available, grouped below by where they belong in an interview.

You can list several techniques on one agent, and should. The agent selects among them based on context: what the respondent just said, where the interview is in the graph, the node's objective. Stacking complementary techniques compounds - active_listening and mirroring reinforce each other because both signal "I heard you" before moving on. Stacking contradictory ones fights itself - putting devils_advocate on a rapport-building opener will make the agent challenge the respondent before any trust is built. See the anti-patterns chapter for the stacking mistakes to avoid.

Separately from technique selection, probing_depth on the agent gates how aggressively any of them fire. light uses techniques sparingly - at most one follow-up, and only when the answer is visibly thin. moderate is the default; the agent probes when it detects a surface answer or a thread worth pulling. deep uses techniques hard - the agent will follow up several times to reach a root cause or a concrete example, up to the max_follow_ups_per_question ceiling. In practice depth matters more than technique selection: a deep interview with a dull technique list beats a light interview with a perfect one.

A technique is not a question. A question is what you ask - it lives in a node in the blueprint graph, written at design time. A technique is how the agent responds to what comes back. You do not pick a technique per node; you pick them once on the agent, and they fire wherever probing is enabled on the graph.

One honest note on evidence. The techniques below are drawn from negotiation, clinical psychology, qualitative research, investigative interviewing, and critical pedagogy. Some have strong empirical backing - active listening has meta-analytic support, and motivational-interviewing reflections have decades of RCT evidence behind them. Others are practitioner craft: widely used, taught in professional training, but lightly studied in controlled settings. Each technique below marks its evidence base honestly. Where a technique is primarily craft, the page says so rather than inventing studies.

Techniques are grouped below by where they belong in an interview, not alphabetically. Jump to a technique:

• Opening and rapport: active_listening, mirroring, labelling
• Going deeper: five_whys, tell_me_more, silent_probe, funnel
• Surfacing the unsaid: socratic, columbo, hypothetical
• Stress-testing claims: devils_advocate, calibrated_questions

Opening and rapport

Active listening

What it is. Reflecting the substance of what the respondent just said back to them, in their own words, before the agent moves on or asks anything new.

Origin. Carl Rogers developed reflective listening as the spine of client-centered therapy in the late 1940s and early 1950s; the Client-Centered Therapy monograph (1951) is the usual reference. The technique crossed from the clinic into interviewing, mediation, and sales training over the next two decades. Weger and colleagues later meta-analysed its use in initial interactions and found a consistent effect on perceived understanding and willingness to disclose.

Why it works. Reflection does two things in one move. First, it confirms to the respondent that they were heard accurately, which reduces the need to repeat or caveat on the next turn. Second, it forces the agent to check its understanding before layering on a follow-up, so if the agent has misheard, the respondent corrects it immediately rather than ten minutes later. The result is denser, more accurate transcript per minute. Weger et al.'s meta-analysis found active listening outperformed simple advice-giving and acknowledgement on every measured dimension of felt understanding.

In action.

Respondent: The pricing jump at our renewal was a surprise - we'd budgeted roughly flat year-on-year and the quote came in almost 40% higher, with no notice.

Agent: So the issue wasn't the price itself so much as the lack of warning - you'd planned for flat and a 40% jump landed without a heads-up. Is that right?

Use when

• The respondent has given a dense or emotionally loaded answer and the agent needs to confirm it heard the claim correctly before probing.
• Early in the interview, when rapport is thin and the respondent needs signal that the agent is paying attention.
• Before any follow-up that assumes a specific interpretation of a previous answer.

Avoid when

• The respondent is already frustrated at being asked to repeat themselves - reflection can read as "you didn't hear me the first time".
• The answer was short and unambiguous. Reflecting "so you use it daily" back after "I use it daily" is filler.

When the respondent shares something substantive, reflect the core of what
they said back in their own words before probing further. Keep reflections
one sentence. Do not summarise multiple points into a tidy package - echo
the specific phrasing that carried the meaning, and invite correction.

Sources: [2], [3].

Mirroring

What it is. Repeating the last one to three words of what the respondent just said, as a question, and then waiting.

Origin. Chris Voss formalised the three-word mirror as a core negotiation move in Never Split the Difference (2016), drawn from his work as the FBI's lead international kidnapping negotiator. The underlying idea - reflecting a fragment of the speaker's language back at them - predates Voss by decades in Rogerian therapy, where paraphrasing and reflective listening cover much the same ground. Voss's contribution was the specific, minimal formulation: last few words, upward inflection, silence after.

Why it works. A mirror adds no content of its own. It signals presence without judgement and without steering, which is unusual enough in conversation that respondents almost always keep talking to fill the gap. Because the agent has not summarised, interpreted, or reframed, there is nothing for the respondent to push back on - so they elaborate in whatever direction feels most true to them, rather than correcting the agent's framing. This is why mirrors tend to surface material the interviewer would not have thought to ask about directly.

In action.

Respondent: Honestly we only ended up renewing because switching felt too painful to take on in Q4.

Agent: Too painful to take on in Q4?

Respondent: Yeah - we'd just hired two new PMs, onboarding them onto a new tool on top of everything else felt insane. It wasn't that your product won, it's that the timing to leave was wrong.

Use when

• The respondent has used a loaded or telling phrase ("too painful", "surprised us", "kind of worked") and the agent wants to surface what sits behind it without leading.
• The agent wants more without committing to a particular follow-up direction - mirroring is maximally open-ended.

Avoid when

• Every turn. Stacked mirrors read as parroting and erode trust fast. Use once or twice per interview, on phrases that genuinely carry weight.
• The respondent is short on time or answering in bullet points - mirroring asks for elaboration the respondent may not have budget for.

A note on stacking: mirroring and active_listening pair well because both signal attention before probing, but do not run both at deep simultaneously. The agent will reflect and then mirror on the same turn, which reads mechanical.

When the respondent uses a phrase that carries weight - emotional language,
an unexpected word, or a hedged claim - repeat the last one to three words
back as a question, and stop. Do not add interpretation. Use this sparingly,
at most once or twice per interview, only on phrases that matter.

Sources: [1], [2].

Labelling

What it is. Naming the emotion or stance behind what the respondent just said, usually with a tentative opener like "it sounds like…" or "it seems like…", and then leaving space for them to confirm or correct.

Origin. Chris Voss, FBI, Never Split the Difference (2016). Labelling sits alongside mirroring as one of the two core moves Voss teaches for getting counterparts to open up under pressure. The tentative framing ("it sounds like", "it seems like") is deliberate - a declarative "you're frustrated" invites denial; a hedged label invites engagement.

Why it works. Two effects combine. First, putting an emotion into words reduces its grip on the conversation - the respondent no longer has to perform around it, because it has been acknowledged. Second, and more importantly for interviewing, a slightly-off label is a better prompt than an open question. A label the respondent agrees with earns a confirmation and often more detail. A label the respondent disagrees with earns a correction - "no, it's more that…" - which is usually sharper and more specific than anything they would volunteer in response to "how do you feel about that?". Either way, the agent learns more than it would from a neutral follow-up.

In action.

Respondent: We'd flagged the data issue three times in the quarterly review and it was still on the roadmap when we left.

Agent: It sounds like what bothered you wasn't the bug itself, it was feeling unheard.

Respondent: Not unheard exactly - we knew they'd logged it. It was more that logging it clearly didn't mean it would get prioritised. We'd lost faith that raising things changed anything.

Note how the slightly-wrong label produced a more precise answer than a direct question about feelings would have. That correction is the value.

Use when

• The respondent has described a situation with clear emotional content but hasn't named the emotion themselves.
• The agent wants a more precise self-report than a direct "how did that feel" would produce.
• Pairing with calibrated_questions later in the interview - a confirmed label is useful context for an open "how" or "what" follow-up.

Avoid when

• The respondent has already named the emotion accurately. Re-labelling a stated emotion sounds like the agent wasn't listening.
• The tone is factual and low-affect - labelling an unemotional answer manufactures drama the respondent didn't bring.

When the respondent describes a situation with clear emotional weight but
has not named the emotion themselves, offer a tentative label: "it sounds
like..." or "it seems like...". Leave space for the respondent to confirm
or correct. A corrected label is often more useful than a confirmed one -
do not steer toward agreement.

Sources: [1].

Going deeper

Five whys

What it is. Following a surface answer with successive "why" (or "why was that") questions, peeling from the stated reason toward the mechanism underneath.

Origin. Taiichi Ohno developed the technique on the Toyota factory floor in the postwar period and formalised it in Toyota Production System (1988). It began as a root-cause tool for manufacturing defects - stop the line, ask why the part failed, ask why that cause occurred, keep going until the chain terminates at something actionable. The move crossed into software incident reviews, UX research, and qualitative interviewing over the following decades, usually losing its original discipline along the way.

Why it works. Surface answers rarely name a mechanism. The first "why" typically returns a rationalisation the respondent has already rehearsed; the second peels to a proximate cause; subsequent turns push toward structural factors the respondent had not yet put into words. The value is not in hitting exactly five turns - the number is indicative - but in refusing to settle for the first plausible-sounding stop.

In action.

Respondent: We stopped using the weekly report.

Agent: Why did the team stop using it?

Respondent: Honestly, it didn't tell us anything the Slack digest wasn't already telling us.

Agent: And why was the Slack digest enough - what did the report try to add that the digest didn't?

Respondent: The report was meant to give week-over-week trends, but by the time it landed Monday morning the team had already moved on to the current week's fires. The comparison was stale the moment we opened it.

Use when

• A respondent offers a tidy single-line reason for a behaviour or decision that the agent has reason to think is downstream of something else.
• The interview objective is a root-cause or diagnostic question - why something failed, why adoption stalled, why a workaround emerged.

Avoid when

• The respondent is speculating rather than reporting. Asking "why" repeatedly about something they did not personally experience just generates confident-sounding guesses.
• The chain has already branched - there is no single "why" when multiple causes interact, and forcing a linear sequence obscures that.

Evidence note: widely used practitioner tool with limited empirical validation. Card (2017) [6] argues the linear "why chain" often stops at a plausible-sounding cause rather than a true root cause - treat output as a starting hypothesis, not a conclusion.

When the respondent gives a single-line reason for a behaviour or outcome,
follow up with "why" (or "why was that") to push past the rationalisation
toward the mechanism. Do not force five turns - stop when the respondent
reaches a description of what actually happened, or when further probing
starts generating speculation rather than memory.

Sources: [5], [6].

Tell me more

What it is. A minimal open prompt - "tell me more about that", "say more", "can you walk me through that" - inviting the respondent to expand on what they just said without specifying which thread.

Origin. The move comes out of the reflective-interviewing tradition. Miller & Rollnick's Motivational Interviewing (2013) is the cleanest reference for the family of "open question plus reflection" moves, of which "tell me more" is the minimal form. It also appears across qualitative-interview methodology - Kvale's InterViews treats the open follow-up as the prototypical probe, the one the interviewer reaches for when the respondent has offered something worth hearing more about but has not been asked to elaborate.

Why it works. The prompt adds no direction of its own. Because the agent has not specified which part of the previous answer to expand, the respondent chooses - and they tend to choose the thread they care most about, which is often not the thread the interviewer would have asked about directly. Every more specific follow-up ("tell me more about the pricing jump" vs. "tell me more") trades some of that signal for focus. Use the minimal form when the respondent has just said something you'd like more of but do not yet know enough to narrow on.

In action.

Respondent: The rollout was fine but onboarding was honestly a bit of a mess.

Agent: Say more about that?

Respondent: The docs were written for a technical admin, but the people actually doing the setup were ops managers. They got to step three, hit an API reference they couldn't read, and opened a ticket. We did that maybe a hundred times in a week.

Use when

• The respondent has offered a loaded summary ("it was a bit of a mess", "we learned a lot", "it kind of worked") and the agent wants the detail behind it without steering which detail.
• Early-to-middle interview, before the agent has a strong enough model to narrow usefully.
• After a reflection or label has confirmed the agent heard the headline - "tell me more" then becomes the follow-through.

Avoid when

• The respondent is time-boxed and answering in bullets. An open invitation to elaborate is the wrong ask if they're trying to keep turns short.
• The thread to pursue is obvious and specific - use a narrower follow-up rather than throwing the choice back to the respondent.

When the respondent offers a loaded or summary-level answer worth
expanding - a phrase like "it was a bit of a mess" or "we learned a
lot" - invite them to expand without naming which thread. "Say more
about that" or "tell me more" is enough. Let the respondent choose
which part to open up.

Sources: [4], [8].

Silent probe

What it is. Deliberately not taking the next turn. The agent acknowledges what the respondent said - minimally, or not at all - and waits, letting the silence do the probing.

Origin. Disciplined silence is a staple of the investigative-interviewing tradition. Shepherd & Griffiths's Investigative Interviewing (2013), the reference for the Conversation Management approach used in UK police and clinical interviewing, treats held silence as a core probe alongside open questions and reflections. Kvale's qualitative-interview literature covers the same ground from the research-methods side: the interviewer's refusal to fill a pause is often what lets the respondent complete the thought they had started to edit.

Why it works. Most people find a multi-second pause in conversation uncomfortable and fill it, usually with whatever they had been about to leave unsaid. The agent's restraint is the probe. Because no new question has been asked, the respondent is not being steered - they are given room to keep going, and what they add is typically the caveat, the second example, or the context they had initially trimmed.

One caveat specific to Debriefer. Silence in a voice interview is embodied - the other person sees you holding the beat. In chat, the respondent only sees the absence of a follow-up. The agent holds by issuing a minimal acknowledgement and waiting for the respondent to type more; treat it as "leave the thread open" rather than a dramatic pause.

In action.

Respondent: The team got behind the decision eventually.

Agent: Mm.

Respondent: Eventually. It took about three weeks, and a fairly painful all-hands where two of the senior engineers pushed back hard on the timeline. I don't think we'd have landed it without that meeting, honestly.

The second paragraph is the value. The agent's minimal acknowledgement and refusal to ask a new question let the respondent complete the thought they had initially rounded off with "eventually".

Use when

• The respondent has given an answer that feels rounded-off or tidy - particularly one that ends with a hedge like "eventually", "more or less", "kind of".
• The agent has a reason to think there's more to come and asking a new question would interrupt the thread.

Avoid when

• The respondent has clearly finished a thought and is waiting for the next question. Holding silence past that point reads as the interview being broken.
• Early in the interview, before any rapport has been built - silence with a stranger is differently loaded from silence with someone you've been talking to for fifteen minutes.

Evidence note: primarily practitioner craft drawn from investigative- and clinical-interviewing training. No direct RCT evidence in research interviewing; endorsed by practitioner literature.

When the respondent ends an answer with a hedge or a rounded-off summary
("eventually", "more or less", "kind of worked"), do not immediately
ask the next question. Acknowledge minimally - a brief "mm" or a short
reflection - and leave the thread open, inviting them to continue rather
than prompting with new content. Use sparingly.

Sources: [4], [9].

Funnel

What it is. Sequencing questions from broad to narrow - starting with an open framing ("walk me through the last time you…") and progressively tightening toward specifics ("what did you click first?", "what did the error say?").

Origin. The broad-to-narrow interview arc predates any single reference, but Kvale & Brinkmann's InterViews (2009) is the cleanest canonical source for it in qualitative research methodology. The principle it names is that the interviewer's narrower questions should be informed by what the respondent said in the broader ones, not imposed ahead of them.

Why it works. A broad opening lets the respondent orient inside their own frame - they choose which event to recall, which actors mattered, which part of the sequence to lead with. That framing is data. Narrowing afterwards lets the agent pull on specific threads without having anchored the respondent on the agent's model of the situation. Running the order in reverse primes the respondent on what the interviewer thinks matters, and they tend to fit their story to that frame whether or not it fits.

One note on where funnel lives in Debriefer. The technique is about structure across the arc of an interview rather than within a single turn: early nodes broad, later nodes narrow. Its natural home is the blueprint graph itself. Setting funnel as a technique on the agent reinforces this at the probing layer - follow-ups open broad and tighten based on what comes back - but if your graph runs narrow-first, no agent-level funnel will recover the lost framing.

In action.

Agent: Walk me through the last time you ran a sales debrief with the team.

Respondent: It was after the Q3 review - we'd lost a big renewal and wanted to go through what had happened. I ran it Thursday afternoon, four people in the room, we worked through the timeline on the whiteboard.

Agent: When you worked through the timeline - what did you actually write down first?

Respondent: The moment the account flagged budget pressure. I remember writing that because everyone agreed afterwards that was when we should have escalated and didn't.

Use when

• The interview objective is a concrete event or process the respondent has lived through and the agent needs both the shape of it and specific details.
• The respondent knows the material better than the agent does, and narrow-first risks encoding the agent's (wrong) model into the question.

Avoid when

• Time is tight and the respondent has already given a clear summary - narrowing directly is more efficient than re-opening.
• The question is genuinely factual and doesn't benefit from framing ("how many people were on the call?" does not need a broad lead-in).

Open follow-ups broad - "walk me through", "tell me about the last time" -
and tighten based on what the respondent offers. Do not jump straight to
a narrow specific question if you have not yet heard the respondent's
framing of the event. Let their broad answer inform which specific thread
to pull on next.

Sources: [4].

Surfacing the unsaid

Socratic

What it is. Treating the respondent as the expert and the agent as someone trying to understand - asking for definitions, distinctions, and the conditions under which a claim would hold, rather than accepting a confident answer at face value.

Origin. The move descends from the Platonic dialogues, where Socrates adopts a posture of not-knowing and asks his interlocutors to spell out what they mean. The modern educational formulation is Paul & Elder's The Thinker's Guide to the Art of Socratic Questioning (2006), which breaks the stance into a small set of repeatable probes: clarification ("what do you mean by that?"), assumption ("what are you taking for granted?"), and consequence ("what would have to be true for that to work?"). Kvale treats the same posture as a standard research-interviewing stance - the interviewer's job is to surface the respondent's reasoning, not to ratify their conclusions.

Why it works. Most respondents hold views they have never had to defend out loud. A confident claim often rides on a definition that has gone unexamined, or on an assumption the respondent would not endorse if it were named. Asking "what do you mean by speed?" or "what would have to be true for that to work?" forces the respondent to articulate the reasoning beneath the headline - which is usually more useful to the interviewer than the headline itself. The technique privileges process over conclusion: how the respondent thinks about the problem is the data.

In action.

Respondent: Our users really care about speed - it's the number one thing driving retention.

Agent: When you say speed, what do you actually mean - page load, time to first value, something else?

Respondent: Honestly, when I say it I mean how fast they can get to an answer. So it's less page load and more whether the report surfaces the number they came for without them having to click through three filters. Which is really a question about defaults, not performance.

Use when

• The respondent has made a confident claim that rides on a term or assumption that has not been unpacked.
• The interview is exploratory or strategic - understanding how the respondent reasons matters as much as what they conclude.
• Pairing with five_whys or hypothetical to push from a stated belief toward the structure beneath it.

Avoid when

• The respondent is time-boxed and answering for recall, not reasoning. Socratic probing asks for reflection, which takes minutes the respondent may not have.
• The claim is genuinely factual and low-stakes - asking someone to define a term they used precisely in context reads as pedantic.

When the respondent makes a confident claim that rests on an undefined
term or an unexamined assumption, ask them to unpack it. "What do you
mean by X?", "what would have to be true for that to work?", or "what
are you taking for granted there?" are the core moves. Stay curious,
not adversarial - you are trying to surface their reasoning, not test it.

Sources: [4], [7].

Columbo

What it is. Signalling that the interview (or a section of it) is over, thanking the respondent, and then coming back with "one more thing…" - a question asked in a deliberately casual register, after the respondent believes the scrutiny has ended.

Origin. Named after Peter Falk's TV detective, whose routine was to close his notebook, turn to leave, and then pivot with "just one more thing, sir" - catching the suspect in a relaxed state. The same move is long-established in investigative interviewing and journalism: the best material often comes after the formal session has wrapped. Shepherd & Griffiths's Investigative Interviewing (2013), the reference for the Conversation Management approach, covers the underlying principle - guarded respondents drop their guard the moment they believe the interview is over, and a well-placed question in that window tends to land differently than the same question asked five minutes earlier.

Why it works. The guard-down effect. Respondents inside a formal interview curate their answers - they think about how each reply will be read, and they edit accordingly. Once they believe they are off-duty, the editing relaxes and they will volunteer the anecdote, caveat, or half-formed thought they had been holding back. The casual register of the "one more thing" reinforces the off-duty frame: the question is signalled as an afterthought rather than a probe, and the respondent answers in kind.

In action.

Agent: That's been really useful - I think that covers what I was hoping to get into. Thanks for the time.

Respondent: No problem, happy to help.

Agent: Actually - one thing before we wrap. You mentioned earlier you almost didn't renew. What was the moment you came closest to pulling the trigger?

Respondent: Honestly? The week the account manager went dark for six days after the pricing call. That was when I started taking the competitor demo seriously. I don't think I ever said that in the QBR.

Use when

• The interview has covered its scheduled ground and there is a specific question the agent suspects would land better outside the formal frame.
• The respondent has been guarded or careful throughout, and the agent has reason to think there is material being held back.

Avoid when

• The interview has genuinely ended and there is no real reason to reopen it - tacking on a question to every interview trains respondents to expect it, and the effect dies.
• The respondent is in a hurry. The move depends on the "off-duty" frame, and someone trying to end a call does not experience a reopened thread as relaxed.

Evidence note: primarily practitioner craft from investigative interviewing and journalism. No direct empirical study of the "one more thing" effect under controlled conditions; endorsed by practitioner literature.

Near the end of the interview, after signalling that the formal session
is wrapping up and thanking the respondent, return with a casual "one
more thing" - a single follow-up on a thread you suspect was held back.
The register matters: ask it as an afterthought, not a probe. Use once
per interview at most.

Sources: [9].

Hypothetical

What it is. Framing a question inside an imagined scene - "if you woke up tomorrow and this was fixed, what would be different?", "if your CEO walked in right now and asked you to justify the budget, what would you say?" - rather than asking for the underlying preference or concern directly.

Origin. Hypothetical and projective framings are a staple of qualitative research. Kvale's InterViews (2009) treats them as a standard move for surfacing preferences respondents struggle to articulate directly. The same family of moves appears in the solution-focused and motivational-interviewing traditions - Miller & Rollnick's work on change talk covers the "miracle question" and related projective framings, which invite the respondent to describe a desired state and in doing so reveal the obstacles they perceive to getting there. The common thread across the traditions is that imagined scenes give the respondent cover to say things they would hedge if asked point-blank.

Why it works. Asking someone what they want, fear, or would pay for often produces a thin, socially-rehearsed answer. The respondent has a version of that answer pre-loaded for the interview and delivers it without much detour through actual introspection. A hypothetical bypasses the pre-loaded answer by putting the respondent inside a scene they have to reason through. They describe what the imagined state looks like, who is in the room, what changes - and the detail they volunteer in the description is where their real priorities sit. The move is useful precisely when the direct question would have been too easy to answer.

In action.

Agent: If you woke up tomorrow and the reporting problem was completely fixed - everything working the way you'd want - what would be different on Monday morning?

Respondent: Honestly, the first thing is I wouldn't open Slack and see the ops team asking whether the numbers are right. That's every Monday right now. And then the exec meeting would be about what the numbers mean rather than whether we trust them. We spend the first twenty minutes of that meeting arguing about methodology.

The answer is sharper than "tell me about the reporting problem" would have produced. The respondent reaches for Monday-morning specifics - Slack, the exec meeting, twenty minutes of methodology arguments - because the hypothetical made them picture it.

Use when

• The respondent is likely to have a rehearsed answer to the direct version of the question and the agent wants to get past it.
• The topic is preferences, priorities, or desired states the respondent has not had to articulate cleanly before.
• Pricing, prioritisation, or change-readiness interviews, where what the respondent would do under a counterfactual is often more revealing than what they say about today.

Avoid when

• The respondent is answering for recall of a specific event. A hypothetical reframes them into speculation and loses the detail of what actually happened.
• The scene you propose is implausible to the respondent. If they cannot picture it, the technique collapses into a stalled question.

Evidence note: supported indirectly by motivational-interviewing research on change talk, but the specific hypothetical-framing move is primarily practitioner craft. No direct RCT evidence.

When a direct question is likely to produce a rehearsed answer, reframe
as a hypothetical - "if this were fixed tomorrow, what would be
different?", "if you had to justify this to your CEO right now, what
would you say?". Put the respondent inside a scene they have to reason
through. Keep the scene plausible to them.

Sources: [4], [8].

Stress-testing claims

Devil's advocate

What it is. Taking the opposing position to whatever the respondent has just claimed - not to change their mind, but to make them argue for their view against a plausible counter rather than simply restate it.

Origin. The role has roots in medieval canonisation proceedings, where the advocatus diaboli was formally appointed to argue against a candidate's sainthood - the church's mechanism for forcing scrutiny on a claim that everyone in the room wanted to believe. The modern empirical treatment comes from Charlan Nemeth's decades of group-reasoning research, summarised in In Defense of Troublemakers (2018). Nemeth's central finding is that dissent improves reasoning quality - but with a sharp caveat: it has to be authentic dissent. A committed devil's advocate is a weak substitute for a real dissenter, though still better than consensus left unchallenged.

Why it works. A respondent asked to defend a claim against a plausible counter-argument has to engage with the counter rather than restate the original claim. That engagement surfaces two useful things. If they have thought through the counter-case, you hear the conditions under which their view would change - the boundaries of the claim. If they have not, the gap is immediately visible: the answer gets vaguer, moves to generalities, or reaches for authority rather than mechanism. Both outcomes are data the direct question would not have produced.

One nuance worth naming. Nemeth's research distinguishes authentic dissent from role-played dissent, and shows the first beats the second on reasoning quality. An agent playing devil's advocate is closer to role-played than authentic - so the technique is genuinely useful for surfacing whether the respondent has prepared for the counter, less useful for actually shifting their conclusion. Treat the output as a test of their preparation, not a test of their view.

In action.

Respondent: The right call is to kill the self-serve tier entirely and go sales-led. The conversion rates are better, the ACVs are higher, and the support burden would drop.

Agent: Let me push back a little - a competitor could read that same data and conclude the opposite, that self-serve is the only way to reach the segment you're currently spending a fortune to acquire through sales. What would you say to that?

Respondent: That's fair on acquisition cost. I think the honest answer is we'd be betting that the segment we can reach self-serve isn't the segment that expands - and I don't actually have strong evidence for that yet. It's more of a prior than a finding.

Use when

• The interview is explicitly stress-testing a claim, decision, or strategy - due diligence, post-mortem, or a pre-commit review where the respondent benefits from having their reasoning challenged.
• The respondent has made a confident claim and the agent wants to know whether the confidence is earned or reflexive.

Avoid when

• The interview is discovery or rapport-led. Pivoting to devil's advocate mid-interview after opening with active_listening and labelling reads as bait-and-switch - reserve it for interviews that are framed as challenge from the outset.
• The respondent is speculating rather than reporting from experience. Arguing against a guess produces a sharper guess, not a truer one.

When the respondent makes a confident claim in a stress-testing context,
offer a plausible counter-argument and invite them to engage with it. Frame
the counter as something a reasonable person might say ("a competitor could
read that the other way", "someone on your board might argue..."), not as
your own view. Push once - do not escalate if the respondent holds the line.

Sources: [10].

Calibrated questions

What it is. Open-ended follow-ups that begin with "how" or "what" - deliberately avoiding "why" and yes/no framings - designed to put the cognitive load on the respondent rather than to elicit a bare confirmation.

Origin. Chris Voss, Never Split the Difference (2016). Voss calls them "calibrated" because each question is shaped to make the respondent do the work: "how can I do that?" forces the respondent to think through constraints on the agent's behalf; "what about this is important to you?" forces them to articulate priorities they had not yet ranked. Voss is specific about the avoidance of "why" - he argues it reads as accusatory in most registers, and pushes respondents into justification rather than description. "How" and "what" carry the same informational weight without the edge.

Why it works. "How" and "what" questions force the respondent to produce an answer that contains structure - a process, a mechanism, a list of constraints, a ranking. Yes/no questions produce thin replies; "why" questions often produce defensive or justificatory ones. Calibrated questions are particularly useful when the respondent has given a confident or closed answer and the agent needs to get them to unpack the reasoning without challenging it directly. The respondent experiences the question as curious rather than adversarial, and answers in kind - but the answer they have to produce is structured enough to be genuinely informative.

In action.

Respondent: We priced the enterprise tier at $50k. It's the right number.

Agent: How did you arrive at fifty specifically - what were you weighing when you landed there?

Respondent: Honestly, it was mostly what the two reference customers had been willing to pay, rounded up a little because we didn't want to leave money on the table on the next deal. We haven't run any willingness-to-pay work on it. It's the right number in the sense that no one has pushed back on it yet, not in the sense that we've tested it.

A "why did you pick fifty?" would have produced a defence. "How did you arrive at it?" produced a process - and the process revealed the number was softer than the opening claim suggested.

Use when

• The respondent has given a confident or closed answer and the agent needs them to unpack the reasoning without reading as a challenge.
• The agent wants structure in the reply - a mechanism, a sequence, a set of constraints - rather than a yes/no or a rationalisation.
• Almost anywhere. calibrated_questions is probably the most broadly useful technique in the list: it pairs with nearly every other technique and rarely fights them. When in doubt, include it.

Avoid when

• A factual, closed question is what you actually need. "How many people were on the call?" is calibrated in form but asking for a number - no need to dress it up.
• The respondent is visibly time-pressed and answering in bullets. Calibrated questions invite structure, which takes turns the respondent may not have.

Phrase follow-ups as open "how" or "what" questions rather than yes/no or
"why" framings. "How did you arrive at that?", "what were you weighing?",
"what would have to change for that to be different?". The goal is to put
the cognitive work on the respondent without reading as adversarial - stay
curious in register, structured in form.

Sources: [1].

References

1. Voss, C. & Raz, T. (2016). Never Split the Difference: Negotiating As If Your Life Depended On It. HarperBusiness. ISBN 978-0062407801.
2. Rogers, C. R. (1951). Client-Centered Therapy: Its Current Practice, Implications, and Theory. Houghton Mifflin.
3. Weger, H., Castle Bell, G., Minei, E. M., & Robinson, M. C. (2014). The relative effectiveness of active listening in initial interactions. International Journal of Listening, 28(1), 13–31. DOI: 10.1080/10904018.2013.813234.
4. Kvale, S. & Brinkmann, S. (2009). InterViews: Learning the Craft of Qualitative Research Interviewing (2nd ed.). SAGE Publications. ISBN 978-0761925422.
5. Ohno, T. (1988). Toyota Production System: Beyond Large-Scale Production. Productivity Press. ISBN 978-0915299140.
6. Card, A. J. (2017). The problem with ‘5 whys’. BMJ Quality & Safety, 26(8), 671–677. DOI: 10.1136/bmjqs-2016-005849.
7. Paul, R. & Elder, L. (2006). The Thinker’s Guide to the Art of Socratic Questioning. Foundation for Critical Thinking. ISBN 978-0944583319.
8. Miller, W. R. & Rollnick, S. (2013). Motivational Interviewing: Helping People Change (3rd ed.). Guilford Press. ISBN 978-1609182274.
9. Shepherd, E. & Griffiths, A. (2013). Investigative Interviewing: The Conversation Management Approach (2nd ed.). Oxford University Press. ISBN 978-0199681891.
10. Nemeth, C. J. (2018). In Defense of Troublemakers: The Power of Dissent in Life and Business. Basic Books. ISBN 978-0465096299.