---
name: interview-and-transcript-analysis
description: >
  Depth analysis of a single interview transcript or a small set, preserving the
  arc of the conversation, what the moderator introduced, and what the participant
  volunteered. Use for "analyse this transcript", "what did this participant
  say", "summarise this interview", "write up these three interviews", "build a
  case study from this interview", "what came out of the depth", "analyse this
  AI-moderated interview", "read this transcript for me".
category: 07 Qualitative Analysis
ref: 07.03
tier: 1
inherits: [K2, K3, K4, K5]
---

# Interview and Transcript Analysis

## 1. One-line description
Analyses a single transcript, or a small set of them, at depth: holding the sequence of the conversation, separating what the participant brought from what the moderator introduced, and preserving the hesitations, contradictions and refusals that carry the meaning.

## 2. What this skill is used for

**The research problem it solves.** A transcript is not a bag of statements. It is a conversation with a shape: an opening where the participant is performing, a middle where they relax, a moment where they correct themselves, a question that put an idea in their head which then reappears twenty minutes later as though it were theirs. Analysis that ignores the shape produces a tidy summary in which the moderator's framing has become a finding, a reconstructed justification has become a reason, and the one moment where the participant hesitated and changed their answer has been smoothed away because it did not fit. This is the default output of both a rushed human write-up and an AI asked to "summarise this interview". The failure is not inaccuracy at the sentence level; each quoted line may be real. The failure is that the account has been detached from the conditions under which it was produced, and those conditions are what tell you how much to believe it.

**Where it sits.** Analysis, immediately after fieldwork. It hands comprehended, annotated transcripts and per-participant accounts to cross-dataset theme building, to case study writing, and to evidence extraction.

**Typical use cases.**
- A single depth interview analysed in its own right, for a case study or a stakeholder account.
- Three to eight interviews from a small qualitative study, analysed as cases and compared.
- Expert or stakeholder interviews where the individual's reasoning is the deliverable, not a theme.
- AI-moderated or asynchronous interviews, where probing was mechanical and needs auditing.
- A pilot batch read before the main analysis, to check whether the guide is working.
- Preparing transcripts for a larger thematic analysis, so cross-participant work starts from comprehension.
- Re-reading a disputed transcript when a finding has been challenged.

**Who uses it.** Qualitative researchers writing up depths; UX and service designers working from a handful of sessions; consultants and strategists analysing expert interviews; academics building case records; anyone who has been handed a transcript and asked what it says.

## 3. When to use it

- The unit of interest is a person, an account or a case, not a pattern across a dataset.
- You have a small number of interviews, typically one to eight, and each one deserves to be understood before anything is aggregated.
- Sequence matters: you need to know what was said before what, and what prompted what.
- The participant is an expert, a decision-maker or a stakeholder whose specific reasoning is the finding.
- The deliverable is a case study, a participant profile, or an annotated account rather than a theme set.
- A larger analysis is coming, and the transcripts need to be understood and quality-assessed first.
- A finding has been challenged and someone needs to go back to the source and check what was actually said.
- The interview was AI-moderated or asynchronous, and the quality of probing needs assessing before the content is trusted.

## 4. When NOT to use it

- **You need themes across a full dataset.** Where the question is what a set of twenty or forty participants collectively show, this skill stops at the case level and **07.01 Thematic Analysis** takes over. The boundary in the other direction: 07.01 explicitly hands single-transcript and small-set work here, because thematic analysis on one transcript produces themes with a base of one, which is a summary with a technical name. **This skill works within and across a small number of interviews at depth, keeping context and sequence. 07.01 builds and tests themes across a full dataset.** Running this skill on thirty transcripts produces thirty case records and no analysis; running 07.01 on three produces three themes that are really three participants.
- **The material is short-form survey text.** A column of one-line open ends has no conversational structure to preserve. Use **07.02 Open-Ended Response Coding**.
- **The transcript is unreliable in a way that cannot be fixed.** Where speaker labels are missing or wrong, where large sections are marked inaudible, or where automated transcription has produced errors that change meaning, coding the text produces confident findings about words nobody said. Assess first (step 1), and where the transcript fails, say which sections are excluded and how much of the interview that removes, per K4 §6.4.
- **The deliverable is a quantified claim.** One interview supports no proportion, and eight support no proportion either. Counting across a handful of cases and reporting it as prevalence is the error that discredits qualitative work fastest. Report cases and their reasoning.
- **You are being asked to confirm what the participant "really meant".** Where a stakeholder has a preferred reading and wants the transcript to yield it, the request is for a search, not an analysis. Say so per K4 §4.2 and offer an explicit assessment of the evidence for and against the stated reading.
- **The interview is in a language or cultural context you cannot read.** Indirectness, deference, humour and refusal are exactly where transcript analysis goes wrong, and translation removes the evidence you would need to catch it. Per K5 §2.2 a reviewer with the relevant context is required before findings are fixed. Where none is available, describe rather than interpret, and say why.
- **Consent does not cover this use, or the participant is identifiable.** A single-case output is inherently identifying, and a case study of one senior person in a small market is a named account whether or not the name appears. This is **13.05 Research Ethics and Consent Design** before it is an analysis question.
- **The recording is the evidence and you only have the text.** Where the finding turns on tone, sarcasm, emotion or a pause, and no audio is available, the transcript cannot settle it. State what the text supports and what it does not.

## 5. Required inputs

**Required.** Without these the skill cannot run. If absent, ask. If no answer is available and the work must proceed, state the assumption at the point where it bites, per K5 §5.

- **The complete transcript, in order, with speaker labels.** Not an extract, not a highlight reel, not a summary produced by someone else. Sequence is the analytical instrument here, and an excerpt has had the sequence removed.
- **A stable participant identifier**, and the identity of the moderator, per K2 §4.2.
- **The discussion guide or question set**, so that prompted content can be told from volunteered content. Without it, every idea in the transcript looks like the participant's.
- **Who the participant is**, in the terms relevant to the study: their role, their relationship to the subject, their recruitment criteria. An interview read without knowing who is speaking is read wrong.

**Optional, and what each one adds.**

- **The audio or video, or timecodes**: permits verification of a disputed passage, resolves inaudible sections, and is the only way to settle a question about tone, sarcasm or hesitation. Without it, the transcript is the ceiling on what can be established.
- **Moderator or fieldwork notes**: record what the room felt like, what the participant said before or after recording, which questions landed badly, and where words and manner diverged. This is context that cannot be recovered from text.
- **The original-language transcript, where the working text is a translation**: allows the translation layer to be assessed rather than assumed away.
- **Other transcripts from the same study**: enables cross-case comparison and reveals whether a striking passage is distinctive to this participant or is something the guide produced in everybody.
- **Anything the participant supplied**: documents, screenshots, artefacts referenced in the conversation. A passage that refers to something you cannot see is only partly readable.
- **A prior interview with the same participant**: makes change over time readable, and makes a reconstructed account easier to detect.

## 6. Questions to ask before starting

1. **Is the case the deliverable, or an input to something larger?** Determines depth and output format. *Default if unanswered:* produce the full case record, which serves both, and note that cross-dataset theme building belongs to **07.01**.
2. **Was the interview moderated by a person, by a script, or by an AI system?** Determines how much probing to expect, how to read a thin answer, and how far the moderator ledger needs to go. *Default:* assume human moderation and check the transcript against that assumption in step 4.
3. **Is audio available?** Determines whether tone, sarcasm and hesitation are evidence or speculation. *Default:* text only, with an explicit statement that paralinguistic readings are not available and any such reading is marked as inference.
4. **Is this a translation, and can the original be consulted?** Determines whether interpretation of wording is permitted at all. *Default:* treat a translation as a second-hand text, restrict close reading of word choice, and flag per K5 §2.2.
5. **What was the participant told this interview was for?** Determines how to read what they emphasised, what they performed, and what they withheld. A supplier interviewed by their client's researcher is managing a relationship as well as answering questions. *Default:* record what is known, and read emphasis with that in mind.
6. **Is this participant likely to be identifiable in the output?** Determines the reporting form before analysis starts, not after. *Default:* assume yes for any single-case output, and flag per K5 §2.4.
7. **How many other interviews exist, and will they be analysed too?** Determines whether cross-case comparison is available and whether a distinctive passage can be tested against anything. *Default:* analyse this case on its own terms and mark what could only be resolved by comparison.

## 7. Step-by-step methodology

**The governing principle.** Everything in a transcript was produced under conditions: who asked, when, after what, in front of whom, and for what purpose. The analysis holds those conditions alongside the content, because the content alone does not tell you how much weight it carries. The order below is not cosmetic. Quality assessment precedes reading, reading precedes coding, and the moderator's contribution is separated before anything is attributed to the participant.

**1. Assess the transcript before you analyse it.** Read for defects, not for content. Look for: missing or swapped speaker labels; sections marked inaudible and how much of the interview they cover; sentences that do not parse; and the characteristic errors of automated transcription, which are homophone substitutions and confident mis-hearings of domain vocabulary rather than obvious gibberish. **The dangerous transcription error is the one that produces a plausible sentence**, because it will be coded, quoted and defended. Where audio exists, spot-check five to ten passages, chosen to include the ones you would most want to quote. *Correct result:* a short transcript quality note stating what is reliable, what is degraded, and what is excluded, with the excluded proportion given. A transcript with 15 percent inaudible is not a transcript with a few gaps; it is a partial record and the analysis says so.

**2. Read the whole transcript before coding anything.** One pass, end to end, without annotating. The purpose is to know how the conversation went before deciding what any part of it means. *Correct result:* a short impression memo, explicitly labelled as impressions, noting where the participant seemed most at ease, where the energy changed, what they returned to unprompted, and anything that surprised you. These are hypotheses about the interview, to be tested against it, not findings.

**3. Map the structure of the conversation.** Segment the transcript into its phases: warm-up, the participant's own framing, the guide's main sections, any deep passage where the conversation left the guide, and the closing. Record where each substantive passage sits. **Position changes meaning.** An answer in the first five minutes is given by someone who does not yet know what the interview is about and is presenting a socially acceptable version of themselves. The same answer forty minutes in, after they have already contradicted it once, is different evidence. A view offered in response to the closing "anything else?" is the thing they were waiting to say, or the thing they thought they should say before leaving. *Correct result:* a structural map, and every later coded passage carries its position.

**4. Build the moderator ledger, and separate the two voices.** Go through the moderator's turns and classify each one: open question, closed question, probe, reflection, summary, leading question, or introduced content. **Introduced content is the critical category**: any concept, word, framing or option that the moderator put into the conversation and the participant did not. Track each introduced term forward. Where it reappears in the participant's speech, mark it. *Correct result:* a ledger of introduced terms with the point of introduction and every subsequent participant use. This is the single highest-value step in transcript analysis, because moderator-introduced content that reappears later reads exactly like a spontaneous participant insight and is routinely reported as one. If a participant says "convenience is the main thing" thirty minutes after the moderator asked "how important is convenience to you?", that is not a finding about convenience. It is a finding about the guide.

**5. Classify every substantive passage on the volunteered-to-prompted scale.** Four levels, and they are not interchangeable as evidence. **Volunteered**: raised by the participant with no prompt on the topic. **Elicited**: given in response to an open question on the topic without the content being supplied. **Prompted**: given after a specific probe naming the thing. **Confirmed**: agreement with a proposition the moderator stated, which is the weakest of the four and is frequently acquiescence rather than agreement. *Correct result:* every passage carries a level, and any passage at "confirmed" is marked as weak evidence in the case record. A theme built on confirmations is a theme built on the guide.

**6. Code within the transcript, keeping sequence attached.** Now code for content, at the level of meaning. Every code instance retains the verbatim span, its position in the conversation, its volunteered-to-prompted level, and whether it followed a probe. Code the participant's own vocabulary rather than translating it into the client's: the word they use for the thing is evidence about how they think about it. *Correct result:* a coded transcript in which any passage can be traced back to where in the conversation it happened and what preceded it. Where this transcript will feed a wider analysis, the codes are provisional and are handed to **07.01** as candidates, not as a frame.

**7. Sweep for hesitation, repair, refusal and absence.** Four separate passes over the transcript, each looking for one thing.

*Hesitation*: pauses, false starts, "I suppose", "I don't know if this is what you want". These usually mark either genuine uncertainty or a point where the participant is deciding how much to say.

*Repair*: the participant corrects themselves. **Record both versions, not the second one.** The move from the first formulation to the second is often the most informative moment in the interview, and summarising it as the corrected version deletes the evidence that a correction happened.

*Refusal*: questions deflected, answered narrowly, answered with a joke, or answered about somebody else. A deflection is data about the topic's sensitivity, and it is not the same as having no view.

*Absence*: things you would expect a person in this position to mention and they did not. Note the expectation explicitly, because an absence is only readable against a stated expectation, and state it as an observation about this interview rather than about the person.

*Correct result:* an annotation set covering all four, with the repairs recorded in both versions. Where audio is not available, hesitation and tone readings are marked as inference, per question 3.

**8. Test the account for reconstruction.** People asked to explain a past decision produce a coherent narrative, and decisions are rarely made coherently. The account you are given has been assembled after the fact, from what the participant remembers, what they have since learned, what makes them look competent, and what fits the story they have told colleagues. **A clean, ordered, well-reasoned account of a messy decision is a signal, not a finding.** Interrogate it: does the sequence they describe match the dates and events elsewhere in the transcript? Do they attribute to a deliberate criterion something that appears elsewhere as an accident? Is anyone else in the account, or have the other people involved disappeared? Do they use the vocabulary of a framework that did not exist at the time? Where reconstruction is likely, report the account as the participant's current explanation of the decision, which is genuinely useful, rather than as the decision process, which it is not. *Correct result:* every retrospective account in the case record is labelled as retrospective, with any internal inconsistency noted, and the distinction between what happened and how it is now explained is visible.

**9. Sweep for within-participant contradiction, and characterise it.** Find the places where the participant said incompatible things. Do not resolve them. For each, record both statements with their positions, what prompted each, and what each was attached to. **Contradiction within one person usually means one of five things**: the two statements are about different contexts that the participant has not distinguished; one is the socially acceptable answer and the other is the lived one; the participant genuinely holds both and has never had to choose; the topic changed between the two and they moved with it; or the moderator changed the framing and the participant followed. Say which reading the transcript supports, or that it supports more than one, at the confidence K3 licenses. *Correct result:* a contradiction register for the case, with each entry characterised rather than reconciled. Per K5 §2.6 this is a review point.

**10. Write the participant-level summary (the case record).** Not a compression of the transcript. A structured account of this person: their situation, how they frame the subject in their own words, their account of the relevant behaviour or decision with its reconstruction caveats, what they volunteered, what needed prompting, where they hesitated or refused, what they contradicted, and what this interview cannot establish. Every claim carries the participant ID and the position in the transcript. *Correct result:* a record from which someone who has not read the transcript could argue with you, because they can see what you built each statement on.

**11. Compare across the small set, case by case.** With two to eight cases, compare cases as wholes rather than pooling their codes. Build a comparison grid: cases as rows, the dimensions that matter as columns, cells filled with the case's position plus a pointer to the passage. Look for four things: dimensions where cases agree; dimensions where they diverge, and what else differs about those cases; a dimension that only appears in one case and may be a person rather than a pattern; and any apparent agreement that turns out to be the guide asking everybody the same leading question. **Do not count.** "Five of eight" on a base of eight is not prevalence, it is a description of these eight, and it will be quoted as a proportion the moment it reaches a slide. *Correct result:* a comparison grid with the reasoning behind each cell traceable, and any cross-case statement written as a statement about these cases.

**12. Produce the case study where the case is the deliverable.** A case study is a claim that this individual account is worth attention in its own right, so it must say why this case, and on what basis it was selected. Structure: who they are in study-relevant terms; the situation; what happened, in sequence; their explanation, marked as their explanation; the tensions in their account; what the case illustrates; and what it does not support. A case selected because it was the most articulate interview is a case selected for eloquence, which is a selection bias with a name (see **07.04**). *Correct result:* a case study that states its own selection rule.

**13. Write the hand-off, and state what remains open.** Name the questions this interview raised that only other interviews or other methods can settle, the passages a human must review, and anything the transcript quality prevented. *Correct result:* an explicit list, not an implication that the case is complete.

## 8. Analytical framework

The chain the output is built on:

    Transcript (quality-assessed) → Structure map → Voice separation → Positioned passage
        → Coded meaning → Annotation (hesitation, repair, refusal, absence)
        → Case record → Cross-case comparison

**Applying it.** The first three steps are conditions, not content, and they are what distinguish this from summarisation. A passage that arrives at the coding stage without its position and its volunteered-to-prompted level has lost the two pieces of information that determine how much it is worth. Everything downstream inherits that loss and nothing recovers it.

**The evidence-strength ladder within an interview.** All four are legitimate evidence and they are not equal.

| Level | What happened | How to read it |
|---|---|---|
| **Volunteered** | Participant raised it unprompted | Strongest evidence of salience. Note where in the conversation it arrived |
| **Elicited** | Open question on the topic, content not supplied | Strong. The topic was the moderator's, the content is the participant's |
| **Prompted** | Specific probe naming the thing | Useful, and evidence about response to the idea, not about salience |
| **Confirmed** | Agreement with a moderator's proposition | Weakest. Frequently acquiescence. Never report alone |

**The three layers of an account.** Distinguish these in every case record, because they are routinely merged.

- **Event**: what the participant says happened. Checkable in principle, against dates, documents or other accounts.
- **Experience**: what it was like. The participant is the only authority and the account is strong evidence.
- **Explanation**: why they say it happened. Retrospective, reconstructed, and evidence of how they now understand it rather than of the mechanism.

Most analytical over-reach in interview work is an explanation being reported as an event.

## 9. Output format

**A. Transcript quality note**
Source and format; whether audio was available; speaker-label integrity; proportion inaudible or excluded; transcription errors found and how; whether the text is a translation and what that limits; the resulting ceiling on what can be established.

**B. Interview map**
Phases in order with approximate positions; where the conversation left the guide; the closing question and what it produced.

**C. Moderator ledger**

| Turn / position | Moderator move | Introduced term or framing | Reappears in participant speech at | Effect on how the later passage reads |
|---|---|---|---|---|

**D. Case record** (one per participant)

    Participant [ID], characteristics, recruitment basis
    Situation: their circumstances in study-relevant terms
    Their framing: how they describe the subject in their own vocabulary
    Account: what they say happened, in sequence, marked event / experience / explanation
    Volunteered: what they raised unprompted, with positions
    Prompted and confirmed: what required a probe, flagged as weaker evidence
    Hesitation and repair: both versions of every correction, with positions
    Refusal and deflection: what was not answered, and how
    Absence: expected content not present, with the expectation stated
    Contradictions: both statements, positions, prompts, and the characterisation
    Reconstruction assessment: which parts of the account are retrospective explanation
    Interpretation: marked as interpretation, with confidence per K3
    What this interview cannot establish

**E. Cross-case comparison grid** (small sets only)

| Dimension | Case 1 | Case 2 | Case 3 | ... | Note |
|---|---|---|---|---|---|

Cells carry the case's position and a pointer to the passage. No counts, no proportions.

**F. Case study** (where the case is the deliverable)
Selection rule stated first; then the structure in step 12.

**G. Open questions and review points**
What only other interviews or other methods can settle; the passages requiring researcher review, per K5 §3.3.

**When the evidence is thin.** A short or unproductive interview produces a short case record, not a padded one. Where a section of the guide was never reached, the case record says so rather than inferring the participant's position from adjacent material. Where the transcript quality caps what can be said, the cap is stated in the case record, not only in the quality note. Where an interview produced nothing usable, that is reported, with the reason, and it is a finding about the fieldwork. A contradiction field reading "none found" is a result; an empty field is a defect.

## 10. Quality checks

Run before anything is presented. Sits on top of K4 §8.

1. Was the transcript quality-assessed before it was analysed, and is the assessment in the output?
2. Was the whole transcript read before any coding began?
3. Does every substantive passage carry its position in the conversation?
4. Does every substantive passage carry a volunteered-to-prompted level?
5. Has every moderator-introduced term been tracked forward, and is any finding resting on a reappearance flagged?
6. Is any content attributed to the participant that the moderator actually said first?
7. Are both versions of every self-correction recorded, rather than only the corrected one?
8. Are contradictions characterised rather than reconciled?
9. Is every retrospective explanation labelled as an explanation rather than reported as an event?
10. Does any statement about absence name the expectation it is measured against?
11. Are paralinguistic readings (tone, sarcasm, hesitation) marked as inference where no audio was available?
12. Is any cross-case statement free of counts and proportions on a base of under ten?
13. Does the case study state why this case was selected?
14. Is the participant identifiable, and if so has that been raised rather than assumed acceptable?
15. Could a colleague who has not read the transcript locate the source of every claim in the case record?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **The moderator's idea returned as a finding** | A striking participant insight uses a word the guide introduced | Step 4. The moderator ledger, every time, and check any headline quote against it |
| **Summarisation dressed as analysis** (the signature AI failure) | The output is a fluent précis in the analyst's register, with the conversation's shape gone | Steps 2 to 5 before any coding. A summary with no positions in it has not been analysed |
| **Sequence stripped** | Quotes appear in thematic order and nobody can tell what came before what | Attach position to every coded span and carry it into the case record |
| **The reconstructed decision reported as the decision** | A clean, rational, sequential account of a messy choice, with no other people in it | Step 8. Label retrospective explanation as explanation |
| **Contradiction resolved toward coherence** | The participant said two incompatible things and only one is in the write-up | Step 9. Record both, characterise, do not reconcile |
| **Repair deleted** | Only the corrected version survives into the record | Record both formulations. The move between them is the evidence |
| **Confirmation counted as agreement** | Findings rest on the participant saying "yes, exactly" to the moderator's proposition | The four-level scale at step 5. Never report a confirmed passage alone |
| **Transcription error quoted with confidence** | A quote contains an odd word that makes just enough sense to survive | Step 1, and spot-check the passages you most want to quote against audio |
| **Counting on a base of six** | "Five of eight participants said" in a small-set output | No counts across cases. Compare cases, describe positions |
| **Eloquence mistaken for representativeness** | The case study is the participant who spoke in complete paragraphs | State the selection rule. See **07.04** on the articulate-respondent bias |
| **Translation read as text** | Close reading of word choice in a translated transcript | Restrict word-level interpretation on translations. Flag per K5 §2.2 |
| **Silence read as agreement** | A topic the participant deflected is reported as settled | Step 7. Deflection is data about the topic, not an answer to it |
| **The interview over-read** | One participant's account becomes a statement about a market | One case supports a case-level claim. Anything wider is a hypothesis, per K3 §4.3 |

## 12. AI guardrails

Skill-specific only. Universal prohibitions are inherited from K4 and are not repeated here. Quote handling follows **K4 §2.3**.

1. **Never attribute to the participant a concept, word or framing the moderator introduced first.** Check the moderator ledger before any finding is written.
2. **Never report the corrected version of a self-correction alone.** Both formulations are recorded, with the sequence.
3. **Never resolve a within-participant contradiction into the more coherent reading.** Record both, characterise the contradiction, and flag per K5 §2.6.
4. **Never present a retrospective explanation as an account of what happened.** Label explanation as explanation, per the three layers in Section 8.
5. **Never infer tone, sarcasm, irony or emotional state from text alone without marking it as inference**, and never at all where the alternative reading would change the finding.
6. **Never report a count or a proportion across a small set of interviews.** Compare cases; do not aggregate them into a number.
7. **Never analyse a transcript without stating its quality**, and never quote a passage from a section marked inaudible or reconstructed.
8. **Never treat a translated transcript as a source for close reading of wording.** Word choice in a translation belongs to the translator.
9. **Never fill a gap in the transcript by inferring what the participant would have said.** A question that was not asked has no answer, per K4 §1.
10. **Never produce a case study without stating why that case was selected.**

## 13. Best-practice principles

1. **Read the whole thing first.** The meaning of minute four is frequently established at minute fifty, and an analyst who codes as they read has already made decisions they cannot see.
2. **Where a thing was said matters as much as what was said.** The first ten minutes are a performance, and everyone performs. The value usually starts when the participant stops explaining themselves to you and starts explaining themselves to themselves.
3. **The moderator is in the data.** Every guide plants ideas. The discipline is not to avoid it, which is impossible, but to track it so that the planting is visible.
4. **The hesitation is the finding more often than the sentence is.** Fluent answers are usually rehearsed ones. The moment somebody struggles to say something is the moment they are working it out rather than retrieving it.
5. **People give you the version of the decision that makes sense now.** Treat every "why" answer as evidence about their current understanding, which is real and useful, and not as evidence about the process, which it usually is not.
6. **A refusal is an answer.** What somebody will not say to a stranger with a recorder running tells you the shape of the topic, and it belongs in the record.
7. **Keep their words.** The vocabulary a participant uses for a thing is evidence about how they categorise it. Translating it into the client's language deletes that before anyone can see it.
8. **One case can carry a strong claim about a mechanism and no claim at all about a population.** Say which you are making.
9. **Compare cases whole, not code by code.** Two participants who used the same word for different reasons will look identical in a pooled code list and completely different side by side.
10. **The interview that felt best is not the most informative one.** Rapport produces fluency, and fluency produces reconstructed narrative. The awkward interview often has more real material in it.
11. **Absence is only readable against an expectation you have written down.** Otherwise it is a gap you noticed after deciding what the finding was.
12. **The analysis is finished when someone could argue with it from the case record alone.** Not when it reads well.

## 14. Worked example

*Fictional scenario, used for illustration only. All participants, quotes and findings below are invented for the purpose of demonstrating method.*

**INPUT.** A business banking provider commissions six depth interviews with finance directors of mid-sized firms that moved their main banking relationship in the last eighteen months. Objective: understand how the switching decision was actually made. The interviews are 60 to 75 minutes, human-moderated, audio available. This example follows one case, P04, a finance director at a manufacturing firm, and its place in the set.

**PROCESS.**

*Step 1.* Transcript quality note: audio available, speaker labels intact, 2 percent inaudible in one passage where two people spoke over each other. Spot-check of six passages against audio finds one substantive transcription error: the transcript reads "we had no relationship manager for a year", the audio says "we had four relationship managers in a year". The corrected version says almost the opposite thing about the bank's staffing and would have supported a different finding. Corrected and flagged.

*Steps 2 to 3.* Whole read. Impression memo notes that P04 becomes noticeably more specific about forty minutes in, after describing a covenant negotiation, and that they return three times unprompted to a single meeting with a credit officer.

*Step 4, and the first judgement call.* The moderator ledger shows the guide introduced the phrase "service quality" at minute 12. P04 uses it at minutes 19, 34 and 58, and by the third use is speaking as though it were their own frame. The draft finding offered was "service quality was the primary driver". The ledger shows the term was the moderator's and every use is downstream of it. What P04 actually volunteered, three times, at minutes 22, 41 and 63, was the specific experience of having to re-explain their business to a new person. The finding is rewritten: **the participant's volunteered concern is repeated re-explanation, and "service quality" is the label the interview supplied for it**. This changes what the client should act on, because the two imply different fixes.

*Steps 5 to 7.* Passage levels assigned. The covenant passage is volunteered. The pricing discussion is entirely prompted and includes two confirmations, so pricing is marked as weak evidence in this case. Repair sweep catches a correction at minute 44: P04 says "it was purely commercial", pauses, then says "no, that's not fair, it was that I'd stopped trusting them to answer the phone". Both versions recorded. The first formulation is the account they would give a colleague; the second is what the interview produced.

*Step 8, and the second judgement call.* P04's account of the switch is orderly: a review was scheduled, criteria were set, three providers were assessed, one was chosen. Tested against the transcript, the sequence does not hold: the "criteria" appear elsewhere in the conversation as things they noticed afterwards, and the review was scheduled two months after the covenant meeting, not before it. Reported as a reconstructed account: the decision as now explained is a procurement process; the decision as evidenced in the transcript began with a single relationship failure and acquired its criteria retrospectively. Both are reported, the difference is named, and the reconstruction is marked at moderate confidence per K3 §4.2, because an alternative reading is that the review was genuinely scheduled and P04 has the dates wrong.

*Steps 9 to 11.* Contradiction register: P04 says the old provider's pricing was uncompetitive and, later, that pricing was "within a rounding error". Characterised as the same-topic-different-context type: the first refers to the headline facility rate, the second to total cost. Not reconciled. Case record written. Cross-case grid across all six cases shows four cases in which the volunteered trigger is a relationship failure and the stated reason is commercial, one where both are commercial, and one where the guide never reached the trigger question. **No count is reported.** The grid statement reads: in five of the six accounts the volunteered material and the stated reason differ in the same direction, which is a pattern in these six cases and a hypothesis for testing at scale, per K3 §4.3.

**OUTPUT.** A transcript quality note with one corrected transcription error that reversed a finding; a moderator ledger showing that the study's apparent lead finding was a guide artefact; a case record for P04 separating event, experience and explanation; a reconstruction assessment; a contradiction register; a six-case comparison grid with no proportions; and a hypothesis, explicitly labelled, that stated switching reasons systematically understate relationship triggers, with the quantitative test named.

## 15. Advanced usage

**AI-moderated and asynchronous interviews.** The moderator ledger becomes more important, not less, because a scripted prober introduces terms with perfect consistency and no ability to notice it has done so. Audit the probe sequence itself: where a script probes on a fixed list, every participant will produce content on every list item and none of it is volunteered. Assess whether the system followed up on the participant's own words or on its own; where it did not, the transcript will look rich and contain almost nothing at the volunteered level. Report the probing quality as a fieldwork finding.

**Expert and stakeholder interviews.** The participant is often managing a position as well as answering. Read for the audience they are speaking to (their board, their sector, their own record), and mark passages where the account serves that audience. Their claims about facts are checkable and should be checked; their claims about their own reasoning are subject to the same reconstruction problem as anyone's.

**Longitudinal re-interviews.** Where the same participant was interviewed before, read both, and treat the difference as the object of analysis. People revise their accounts of stable events, and the revision is usually more informative than either version. Do not correct the earlier account with the later one.

**Disputed findings.** Where a stakeholder challenges a finding, go back to the transcript with the challenge stated as a proposition, and assess the evidence for and against it explicitly rather than re-reading until the original finding is confirmed. Record the challenge before re-reading, so confirmation is visible if it happens.

**Feeding the wider analysis.** Hand case records, not raw transcripts, to **07.01**, along with the moderator ledger for each interview. Theme building that starts from comprehended cases is faster and much less likely to build a theme out of the guide. Hand annotated passages with positions and levels to **07.04** for evidence selection.

**When the standard approach does not fit.** Where the interest is the language itself rather than what it reports, this is discourse or narrative analysis, which this skill does not cover. Where the interview is one of many dozens, run this skill on a purposively selected subset for depth and **07.01** across the rest, and say which is which.

## 16. Skill chain

**Recommended previous skills:**
- **02.02 Discussion Guide Design**, retrospectively, to establish what was asked, in what order, and what the guide introduced.
- **03.02 AI-Moderated Interview Design**, where the interview was machine-moderated and the probing logic needs to be read alongside the transcript.
- **13.05 Research Ethics and Consent Design**, where a single-case output would identify the participant.

**Recommended next skills:**
- **07.01 Thematic Analysis.** Takes case records, moderator ledgers and provisional codes, and builds tested themes across the full dataset.
- **07.04 Quote and Evidence Extraction.** Takes positioned, level-classified passages and selects verified evidence for the report.
- **08.01 Finding to Insight Development.** Takes case-level accounts with their reconstruction caveats intact.

**Runs well alongside:**
- **07.06 Qualitative and Quantitative Integration**, where a case-level mechanism needs testing against a survey.
- **13.03 AI Output Verification**, run against the case record and any quote leaving the analysis.
- **K5**, at the three review points this skill mandates: ambiguous or contradictory passages, cultural and linguistic context, and any output in which the participant is identifiable.

---
A Yazi Supplied Skill and resource.