# An oral history collection — interviews summarised, indexed and released with consent Recipe No. 43, Research and civic life. From The know.sh Cookbook: https://know.sh/cookbook/oral-history - For: a volunteer at a town historical society cataloguing an oral history project - You bring: recordings already held in the society’s archive, signed release forms, the interviewers’ field notes, time to listen to each interview once, closely, and the AI assistant you already use - You get: a collection with a document per interview, time-coded summaries, short key excerpts, an index of names and places across every interview, and public links for the interviews the narrators have released - Time: about twice an interview’s length to listen and take notes, then half an hour to check what your assistant filed - Keep it: private while you work; share one document by a read-only link The Ellery Falls Historical Society has thirty-one recorded interviews: quarrymen, mill workers, a schoolteacher, the last station agent. The recordings are safe in the archive. What nobody can do is find anything in them. A researcher who asks "did anyone talk about the 1955 flood?" gets a shrug and a list of names. This recipe catalogues the collection. Each interview becomes a research document; its findings are time-coded segments summarising what the narrator talks about, a handful of short excerpts transcribed word for word, the names and places mentioned, and the release status. You do the listening. Your assistant — Claude, ChatGPT or a local model — turns your listening notes into the documents and findings, and you check them against the recording in the editor. The shelf's index then does what the shrug could not: *Hollis* leads to every interview and segment where the mill or its owner comes up. For interviews whose narrators have signed a release, a public link lets the town read the summary, with any restricted segment left out. The recordings themselves never leave the society's archive. know.sh holds the catalogue that points to them. ## What you will use - **Shelf**: One shelf for the project, *Quarry and mill voices*, with the collection’s archive reference in its line. - **Research document**: One document per interview; the overview gives the narrator, date, place, interviewer, length and recording number. - **Finding**: Time-coded segments as *Observation*, key excerpts as *Evidence*, the release as a *Decision*, and things to follow up as *Question*. - **The A–Z index**: The index of names and places across every interview, built from segment titles and first lines, which is the finding aid the collection lacked. - **Your AI assistant**: Turns your listening notes into an interview document with its segments, reads across interviews for dates and names that disagree, and drafts follow-up questions. - **The editor**: Where you check each time code and summary against the recording, type the excerpts yourself and set the restricted segments apart. - **Proposed notes**: Your assistant’s proposed notes where one narrator’s date or spelling differs from another’s, for you to accept or dismiss. - **Public link**: A public link per released interview, with restricted segments and the release record left out. ## Method ### 1. Set up the collection Decide the shelf's name and give it the archive reference as its line: "Quarry and mill voices, oral history collection OH-2025, Ellery Falls Historical Society." In the shelf notes, write down the project's standards: how segments are timed, how names are spelled, what the release form allows. Connect your assistant to know.sh once, following the support page, and ask it to make that shelf. Each interview will be one document, titled with its number, the narrator and the subject: *Interview 7 — Marjorie Tate: the quarry and the 1971 closing*. Its overview is the catalogue record: narrator, date and place, interviewer, length, the recording's archive number, and one line on its release status. ### 2. Listen, and give your assistant the notes Listen with the recording open in the archive's player and type notes as you go. Every time the conversation turns, write the time code, the subject, the people and places, and a line or two on what the narrator says. Eight to fifteen segments for an hour's interview is usual. For a restricted segment, write only its time codes and the word *restricted*. Give those notes and the catalogue record to your assistant and ask it to make the interview's document: the record as the overview, and one *Observation* finding per segment, titled with its time code and subject — "00:14:20 — The quarry whistle and Quarry Road" — with a third-person summary drawn only from your notes. The recording itself does not go to the assistant. ### 3. Check the segments and transcribe the excerpts yourself Open the document in **Edit** with the recording beside you. Check every time code and every summary: the assistant never heard the interview, so anything it smoothed over or guessed is yours to correct. Then add the passages that deserve the narrator's own words: three to six short excerpts, transcribed word for word by you, each a finding filed as *Evidence*, titled with its time code and subject. Keep them short. The recording is the source; the excerpt is a pointer to it, and a researcher who wants more comes to the archive. ### 4. Record the release and any restrictions Give every interview a *Decision* finding titled "Release and restrictions". Record the date the release was signed, what it permits, where the signed form is kept, and any restriction the narrator asked for: "00:41:10–00:47:30 closed until 2040 at the narrator's request." A restricted segment is a finding of its own, holding its time code, a note that it is closed, and the restriction; no summary of what it says. That way it can be left off a public link without cutting anything else, and nobody reading the catalogue learns what the narrator chose to hold back. ### 5. Let the index become the finding aid Open the shelf's *Index to this shelf*. The people and places named in segment titles should appear with the interviews and segments where they come up: *Tate*, *Hollis*, *Ellery*, *Congregational*. A name that appears only deep in a summary may not reach the index, so if a researcher would look for it, put it in the segment's title. Agree one spelling for each name and place, and write it in the shelf notes. When narrators disagree — the Hollis mill or Hollis's mill — pick one form for titles and keep the other in the summary. ### 6. Ask your assistant to read across the interviews Ask your assistant to compare interviews on the shelf: which narrators mention the 1971 closing, and do they agree on the month? It can leave a proposed note, with **Accept** and **Dismiss**, wherever one interview's date or spelling differs from another's; accept the ones you will act on and dismiss the rest. Ask it, too, for follow-up questions before an interviewer goes back to a narrator. It works only from your summaries and excerpts, not the recordings, so check anything it suggests against what you heard. ### 7. Publish released interviews, one link each For each interview whose narrator has signed a release, press **Share** and make a link with no expiry. Leave out the *Release and restrictions* finding and every restricted segment from each finding's page. Readers see the overview, the segments and the excerpts, without your notes or highlights. Copy the link when it appears; it is shown once. Make a *Read me first* document listing the released interviews and their links, share that, and put its link on the society's website. Interviews without a signed release get no link. ## Prompts to try Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > Here are my listening notes and the catalogue record for Interview 7. Using know.sh, create a document titled “Interview 7 — Marjorie Tate: the quarry and the 1971 closing” on my shelf “Quarry and mill voices”. Make the catalogue record its overview, and add one Observation finding per segment in my notes, titled “time code — subject”, with a short third-person summary taken only from my notes. For segments marked restricted, add a finding with the time codes and “Closed at the narrator’s request” and nothing else. Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > Using know.sh, read every interview on my shelf “Quarry and mill voices” and list each event that more than one narrator describes (the 1955 flood, the 1971 closing, the station fire), with the date or detail each gives and the interview and segment numbers. Where narrators disagree, leave a proposed note on each finding involved. Do not change the summaries. Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > Using know.sh, check every document on my shelf “Quarry and mill voices”: list any interview without a “Release and restrictions” finding, and any segment whose time code is out of order. Then draft ten follow-up questions for a second visit to the narrator of Interview 7, grouped by segment. Put the questions in your reply, not in the library. ## Variations - For a family rather than a town, record grandparents the same way and keep it private; the [family history](/cookbook/family-history) recipe shows how the interviews feed a family book. - Make a walking tour from the collection: one document, a finding per stop, each quoting a short released excerpt and linking to the interview it comes from. - Using interviews as evidence in a dissertation? The [thesis research base](/cookbook/thesis-research) files them under the chapters whose arguments they support. ## Where it falls short - know.sh cannot hold or play audio, and nothing is uploaded; the recordings stay in the society’s archive and each document names its recording number. - Your assistant never hears the recordings in this recipe. Every summary rests on a volunteer’s notes, and every excerpt on a volunteer’s transcription. - Smaller local models call tools less reliably and can drop a segment from a long batch. Check each document against your notes; Revisions shows every change. - Links are one interview each, and a whole collection cannot be shared as one. A *Read me first* document with the links is the way round it. - This is a catalogue, not an archival management system. Accession records and the preservation copies of the recordings belong in the society’s archive systems. ## A note on consent, releases and the people in the stories Follow the Oral History Association's principles and best practices: informed consent before the interview, a signed release before anything is made public, and the narrator's right to restrict all or part of it. Publish only interviews with a signed release, and honour every restriction it records. When in doubt, ask the narrator again. Narrators talk about other people, many still living. Before sharing an interview, read it for private details about third parties — health, family quarrels, accusations — and leave those segments out of the link. A summary can be defamatory even when the recording sits safely in an archive. Releases rarely say anything about AI services. Give your assistant your own notes, not the recordings or full transcripts, and never the content of a restricted segment, unless the society has decided its releases allow it. Assistants make mistakes, and so does memory. Check dates and names against newspapers and records before they appear in anything the society publishes. Indexed under: Oral history, Release forms, Time codes, Finding aids, Restricted segments, Historical societies, Interviews.