Family · Guide №0003 · 4 min read
Ask what they witnessed, not what they lived.
ChatGPT's free voice mode will interview your relatives for you — one question at a time, up a 12-rung ladder from facts to feelings — no card, no interviewing skill. One setup prompt, one branch chosen by their temperament, one session this week. The catches, stated up front: the free tier's daily voice allowance is unpublished and varies (a session can cut out mid-rung), chats train OpenAI's models unless you flip one toggle first, and the transcript it saves you is not the same artifact as the voice.
Numbers wearing an "as of" chip drift. The ladder, the reframe, and the branches don't.
The centerpiece — Pick the Branch
The interview forks on the subject, not the topic
Answer two questions about the actual person. Each path lands on a different session plan — different setup, different first question, different steering rule. Nothing is sent anywhere; this runs entirely inside the page.
QUESTION 1 OF 2
The one-line reframe
Ask what they witnessed, not what they lived.
Protagonists freeze. Witnesses talk.
01 — Why the ladder wins
An interviewer, not a compiler
Every generic version of this workflow ends the same way: paste the recording into an AI and ask for a memoir. That ending destroys the one thing the process actually produced — her voice, her pauses, her laugh in the middle of the sad story. The compiled book is the least valuable artifact; the recording is the heirloom. So here the AI works at the front of the process, conducting, and the raw session is the deliverable.
The 12-rung ladder is the engine. "What's your biggest regret" asked cold gets a deflection; asked forty-five minutes after the smell of her mother's kitchen on Sundays, it gets the answer. Depth isn't requested — it's earned by rungs 1–9. Rung-10 answers don't exist on visit one, which is why question lists fail: they hand you rung 10 on page one.
And the witness reframe is what unlocks the subjects who insist nothing interesting ever happened to them. "What did you do" demands a performance — a life summarized, judged. "What did you see" offers company — the subject becomes the storyteller about everyone else, and their own story slips in sideways, unguarded.
When a session matters enough for permanence beyond your family, StoryCorps (free) is the institutional deposit — the Library of Congress keeps those recordings. Pair it with this protocol: the ladder runs the interview; StoryCorps keeps it forever.
What it replaces
Tick what you'd otherwise buy to capture the stories — then run the swap.
List prices, secondary-source — treat as estimates, not quotes.
02 — The walkthrough
Five steps, one archive
Flip one switch before anything else
Create the free account (email, no card — T0) and, before the first session, open Settings → Data Controls and turn off "Improve the model for everyone." Free-tier conversations are training data by default: names, birthdates, and the story about the uncle nobody mentions.
There is no un-telling a story to a training corpus. The toggle takes forty seconds and belongs before session one, not after.
Write the interviewer's instructions
Open a chat, switch to voice mode, and paste this before the subject arrives:
You are conducting an oral-history interview with my [relation], [name], born [year]. Your rules: • Ask exactly ONE question at a time, then stop and wait — never fill silence, however long. • Follow this ladder, advancing one rung only after a full answer: rungs 1–3 FACTS (streets, jobs, prices, neighbors) · rungs 4–6 SENSES (smells, sounds, objects) · rungs 7–9 PEOPLE (who argued, who laughed, who disappeared) · rungs 10–12 MEANING (proudest moment, biggest regret, advice for a great-grandchild). • Build follow-ups from [name]'s own words. • Never correct, never rush, never judge — you are a witness, not an editor. Begin with the question I give you after this message.
The silence rule is in the prompt on purpose: the AI will wait even when the humans in the room can't. That's the machine's one unbeatable advantage over a nervous grandchild.
Take the branch the subject dictates
Run Pick the Branch at the top of this page — the interview forks on them. The short version: freezers get object-anchored questions ("before you tell me who's in this photo — where was this kitchen, and what did it smell like on a Sunday?"), ramblers get one-story-at-a-time steering, heritage-language subjects get the whole interview in their strongest language — memories encoded in a first language surface in that language — and the tech-averse get a printed deck with AI doing the before (questions) and after (transcript cleanup) instead.
Getting this branch wrong isn't fatal, but it costs you a session: an object question to a natural storyteller wastes their momentum, and a life question to a freezer ends the interview in ninety seconds.
Run the session, and out-sit the machine
Phone flat on the table between you, voice mode on, speaker up — and a second device recording the room on plain Voice Memos. That's dual capture: the transcript is for searching, the audio is the artifact. Twenty to forty minutes; the free voice ceiling is more likely to spare you at that lengthas of 2025-06-26.
Your only job is the hard one: when she goes quiet after a hard question, nobody speaks. The ten seconds of silence is where the real material surfaces — every helpful fill cuts off a story that was fifteen seconds from arriving.
Close the loop between sessions
The archive compounds only if each session starts where the last ended. Paste the transcript into a fresh chat and send:
Here is the transcript of session [n] with [name]: [paste transcript]. Do three things: 1. List the three threads [name] left unfinished — quote the exact lines. 2. Write the next session's questions for rungs [x]–[y], each one built from a word [name] actually used. 3. Flag any names, dates, or places I should verify against family records. Do not invent details that are not in this transcript.
The last line is load-bearing: chat memory drifts, and an AI that "remembers" a detail that was never said will build follow-up questions on top of it. Paste the transcript; trust nothing else. For sessions worth permanence, StoryCorps (free) deposits recordings with the Library of Congress.
03 — Archive math
What does the plan actually capture?
Framework constant: the full ladder is 12 rungs × 3 questions × ~7 minutes ≈ 252 minutes of interview. Ceiling: one session per day is the practical max on free voiceas of 2025-06-26 — the daily limit is unpublished and varies.
What breaks
The silence is the answer working. The subject stares at the table, someone says "you don't have to answer that," and the story that was fifteen seconds from arriving never does. The AI waits because it was told to. Your job is to wait longer than the machine.
- Voice caps are unpublished and variable. The free tier's daily voice allowance has no published number — a countdown appears only when you're near it, and a long session can end mid-rung. The protocol survives it: the continuity prompt restarts any cut session from its transcript in a fresh chat. Keep sessions at 20–40 minutes and you'll rarely see the ceiling.
- Training on by default. Free-tier chats improve the model unless you opt out. Flip the Data Controls toggle before session one — family names and dates in a training corpus can't be recalled.
- Transcript ≠ audio. The shared-link transcript strips the pause before the hard sentence and the laugh in the middle of the sad story. Dual capture is the difference between an index and an heirloom.
- Invented follow-ups. Asked to "continue last time," a model will sometimes recall details that were never said — and build questions on them. Always paste the transcript; treat any follow-up that doesn't quote a real line as suspect.
- Accents and dialects. Heavy accents garble transcripts. The audio is the truth; the transcript is the finding aid — label it that way in your archive.
- The deflector. "Nothing interesting ever happened to me" is near-universal, and the witness reframe exists for it — but if three object-anchored questions still get nothing, drop to rung-1 facts and let the ladder work. Pushing feelings on visit one buys you no visit two.
- Grief timing. After a loss, the recordings you have become fixed and the ones you planned become the regret. One imperfect session this month outranks a perfect protocol next year — the phone call to book it is the hardest step in the entire method.
05 — Three things worth keeping
If you forget everything else
Depth is earned by escalation, not requested. A question list hands you rung 10 on page one and gets a deflection; forty-five minutes of streets, smells, and neighbors gets the same question answered without being asked.
"What did you do" demands a performance; "what did you see" offers company. The subject becomes the storyteller about everyone else — and their own story slips in sideways, unguarded. This one reframe converts most "nothing interesting" subjects into talkers.
The compiled memoir is the least valuable output — it strips exactly the part you can't get back. Record the room on a second device, keep the transcript as the finding aid, and let the AI's job be the interviewing, not the rewriting.
06 — Try it now
This week, not "sometime"
- Pick the relative whose stories you'd miss most.
- Flip the training toggle (Settings → Data Controls) — forty seconds.
- Run Pick the Branch for them; copy their path's opener.
- Paste the interviewer prompt; start the second-device recording.
- Book twenty minutes. The phone call to book it is the hardest step in the whole method.