An interview that isn't put on the record promptly gets misremembered — confirmation bias quietly keeps the parts the interviewer liked and drops the rest. The original method keeps a spreadsheet: one row per interview question, one column per interview, notes in the cells. This skill is that spreadsheet as files: after each conversation it distills whatever raw material exists — a transcript, hasty notes, a voice-memo dump — into one brief, structured debrief file, mapped question by question, exact words preserved, gaps honest, surprises kept even when they fit nowhere. The user walks away with evidence a later synthesis can use without ever re-reading the transcript.

The interview method chains numbered artifacts: goal questions (G1, G2, …) that the user needs answered but can't ask directly; hypotheses (H1, H2, …), the user's falsifiable best guesses, mapped to goals; and open-ended interview questions (Q1, Q2, …), each a miniature experiment mapped to the hypotheses it tests. A debrief completes the chain for one conversation: each answer sits next to the question it answers, which carries the hypotheses it tests. Because every debrief uses the same Q-numbers, answers become comparable across interviews — that's what lets a later pass see that five of seven people said the same thing. Unmapped notes are anecdotes; mapped notes are evidence.

A debrief is not a summary essay; it's a row of cells. Each answer is one to three lines: the fact, the number, the story in miniature — with the customer's key phrases quoted verbatim. The reader of this file is a future synthesis pass across many debriefs; ten pages per interview would defeat it. Compression is honest as long as nothing is added: shrink the words, never the meaning.

Some of the most valuable findings are vocabulary: what customers call themselves, their work, the product category, the pain. Positioning gets built from customers' exact words, not the company's. So when a phrase does work — how they describe the problem, what they call the tool, an emotionally loaded word they volunteered — it goes in the file in quotation marks, verbatim, even inside an otherwise compressed answer. A paraphrase of vocabulary is destroyed vocabulary.

Conversations stray from the script, and straying is often where the new learning is — attitudes and behaviors the hypothesis list never anticipated. Everything interesting that fits no question goes in an addenda section: verbatim where the words matter, one bullet per observation. Never discard something because there's no slot for it; the method expects the best discoveries to arrive slotless. Slotless doesn't mean unmapped, though: when an addenda item plainly bears on a hypothesis, tag it with the [H-numbers] — evidence is worth more mapped. Mark the items that genuinely surprised the interviewer — surprise is the signal the whole method runs on, and the later synthesis step mines it.

Nearly every interviewee at some point stops reporting their own life and starts speaking for the market: "I'd pay $50, but most people would expect this free." People are experts on themselves, not on everyone else — that's why interviews exist. Record such an answer, but flag it as market-guru commentary; where it contains a real self-report ("I'd pay $50"), extract that part as the actual answer. The flag lives wherever the moment occurred — inside the answer cell if it arrived within a question, as an addenda bullet if it was freestanding. The flag matters later: guru claims weigh almost nothing as evidence about the market, full weight as evidence about the speaker.

Write to be understood, not admired. The work here wrestles with hard concepts, and clever metaphors, wordplay, or cute turns of phrase make them harder to grasp, not easier. Say plainly what you mean. If a sentence reads more clearly without a flourish, cut the flourish. State the actual point rather than gesturing wittily at it.

When you mention a numbered or lettered item to the user — K4, W2, O17, H3, and the like — add a few plain words on what it actually is ("K4 — the owner whose career rides on the site"). A bare token is unreadable to a human who saw it defined hours or days ago: the tag is for traceability, the gloss is for comprehension. Keep the tag for accuracy; always add the gloss.

Every answer in the file traces to something the customer actually said (or, for reconstructed notes, something the interviewer attests they said). Never infer an answer to a question that wasn't asked, never extrapolate from an adjacent answer, never smooth a rambling answer into a cleaner position than the customer took. Ambivalence, contradiction, and "they didn't really answer" are all legitimate contents of a cell. If the user's recollection conflicts with the transcript, show both and ask — transcripts contain errors, and memories contain wishes; neither wins automatically.

The file is the customer's testimony, lightly annotated. Commentary is allowed — one line per answer at most, clearly marked — and only when it captures something available in the moment that the future analysis would otherwise lose: tone ("visibly angry retelling this"), context ("note: 10× bigger than our target segment"), a flag ("directly contradicts what we assumed"). Commentary may also carry the user's own marked notes — context, tone, an untested expectation ("interviewer expects a lost-job trigger would move him; untested") — and factual cross-references to a sibling debrief ("business partner, interviewed same day, gave a conflicting ceiling"). Flags, never conclusions: commentary never argues, never editorializes at length, and never proposes hypothesis changes — that is the next step's job, done across many debriefs, not inside one.

The paper is not the whole interview. After drafting from the raw material, turn to the user for what the paper can't supply: garbled or ambiguous passages ("did you read this as X or Y?"), answers the notes compress past usefulness, and the standing question — "did anything surprise you that isn't in these notes?" Also distinguish two kinds of blank: a question never asked versus asked-but-not-written-down; the user knows which, and the file should say. Ask in small batches, a few related clarifications per exchange, not a wall of twenty. Whatever arrives this way goes in marked — a cell recovered from the interviewer's memory says so inline ("from interviewer's memory; asked after the recorder stopped") — so a mixed-provenance file stays honest cell by cell.

If the raw material covers several conversations, split it and process them one at a time, each into its own file. Mixed debriefs destroy the one-column-per-interview property that makes patterns visible.

Gather, asking only for what's missing:

Where the file goes: an directory next to the question list (create it if absent; if the question list isn't on disk, ask where the method's files live — default: the current directory — and put there), one file per conversation, named . If a debrief for this conversation already exists, ask whether to revise it rather than silently writing a duplicate.

Walk the raw material against every question in the list, in question order:

Then sweep again for everything interesting that fit no question — the addenda: off-script stories, unprompted vocabulary, coping mechanisms, buying context, emotional reactions. Mark items that seem to have surprised the interviewer — in-transcript cues (an audible "wow," an abandoned question order) are candidates to confirm at the clarification pass. Note follow-up threads the interviewer chased and what they yielded; those usually land in either the parent question's answer or the addenda.

Show the user the complete draft debrief — this is a review of an artifact, not a co-forged list, so presenting it whole is right. Then run the clarifications, a few related items per exchange:

Fold in corrections. Where the user's correction contradicts the transcript rather than supplementing it, show the discrepancy and let them decide — but record which source the final cell came from.

<Two to four lines: who this was, how the conversation came about, and any circumstance that colors the answers ("was between jobs, kept it to 25 minutes"). Provenance of this record: transcript / notes / reconstructed from memory.>

Maps to: <question-list file, e.g. ../QUESTIONS.md>

Q1. <Brief answer; "their key phrases" verbatim.> [H1, H4] 💬 <optional one-line interviewer commentary>

Q2. — not asked

Q3. Asked; no real answer — <dodged / didn't know / out of time>. [H2]