How to Build a Qualitative Theme Codebook from Notes
Interview codebooks fail when the model writes the field. It invents a memorable quote, a participant named Maya, and a saturation claim. You present themes. A co-coder asks which transcript. There is no transcript line. The job is a theme codebook from the notes you pasted, with names stripped, and no minted speech.
The matching generator is the Qualitative Theme Codebook from Interview Notes prompt. Browse related cards in the PromptDig library (Browse more prompts). When a filled scheme survives a check against the notes, share the version you actually use (Share a prompt).
De-identify the notes before you name a theme
Question is the analytic job: inductive themes, a focused codebook, or a map onto an existing framework you pasted. Sources: protocol, prior codebook, inclusion criteria. Pastes: interview notes, memos, and any already-approved theme list. Source policy: notes only. No "what people usually say about onboarding." Output shape: theme names, definitions, inclusion and exclusion, and excerpts only from Pastes. Off-limits: invented quotes, real names, workplaces, schools.
The first pass is a scrub list. Emails, phones, children's names, employer names, and rare job titles that identify a person come out. Replace with roles if the notes already use roles. If the notes still have a unique story that identifies someone, flag it. Do not publish that excerpt.
Themes may be named only from patterns that are in the notes. If five notes mention wait time and two mention price, you may propose Wait time and Price as candidate themes. You may not add Trust in leadership because it is a popular org-psych theme.
No invented quotes, saturation, or n
Excerpts: only text that is in Pastes, shortened if needed without changing meaning, and tagged to a de-identified ID that is in the notes (P3, Site B). Do not invent P7. Do not clean a paraphrase into quotation marks.
Counts: only if you counted in Inputs. Do not print 7 of 12. Do not print saturation. Saturation is a methodological claim you did not paste. If the notes are six bullet pages from one day, say six pages from one day.
Definitions: observable, not poetic. "Frustration with wait time" needs a rule: mentions delay, calendar, or queue. "Culture of care" is not observable unless the notes use that phrase. Prefer the participants' words when those words are in the paste.
If a framework is in Sources (COM-B, a prior codebook), map only where the notes support the map. Unmapped notes stay Unmapped, not forced. Do not invent a construct to look complete.
Method, disagreements, and a book two people can use
Method notes: how notes were taken (live, after, AI transcript you pasted), how many interviews the notes cover if that number is in Inputs, language, and whether you are open or closed coding. If the count of interviews is missing, write unknown. Do not infer n from filenames.
Disagreement rule: what happens when two coders split. If you did not specify, list it as a gap. Do not invent a kappa target.
Gaps: missing audio, missing consent language, missing whether notes are verbatim, missing setting, missing whether the study is IRB-covered. Print those. This is research-adjacent. Add not-advice. It is not a findings report and not a journal article.
Banned words: best, guaranteed, #1, invented percents. "The majority felt unheard" is a results claim. A codebook should not make it.
A worked pass: notes with a Friday deadline
Say Pastes has messy notes from two conversations, a Friday deadline, and no approved quote list. Budget not approved for transcription. The codebook can list candidate themes from repeated bullets in those notes. It cannot open with a pull-quote. The NO_DATA list should include verbatim excerpts if the notes are paraphrased, participant n if missing, and any theme that would require a third interview you do not have.
If names leaked into the notes, the scrub list should show what was removed. If the generator still writes Maya from Site C, that is a miss. Rerun after you strip.
A second miss is collapsing disagreement into one theme to look tidy. If one note says the tool is slow and another says training was missing, those can stay two themes. Tidiness is not validity.
When you later get transcripts, paste de-identified excerpts and rerun. Do not keep invented quotes from the notes-only run. The source policy echo should still say notes only, plus the new transcript paste.
Fill the card, then run
Paste de-identified notes, the protocol question, and any prior codebook. Leave Off-limits explicit.
Question: [Question]
Sources I may use: [Sources]
Pastes / files: [Pastes]
Source policy: [Policy]
Output shape: [Shape]
Off-limits: [OffLimits]
A theme book is a shared lens, not a highlight reel. When a run refuses to mint a quote because the notes were paraphrased, share that filled card.