How to Build a Survey Open-End Codebook from a CSV

Open-end codebooks fail when the model writes the answers. It invents three vivid quotes, a 42% theme share, and a code that no respondent used. You brief the team. Someone asks which row said that. There is no row. The job is a codebook from the CSV description and sample you actually have, with definitions tied to columns, and no minted response text.

The matching generator is the Survey Open-End Codebook from a CSV Description prompt. Browse related cards in the PromptDig library (Browse more prompts). When a filled codebook survives a spot-check against real rows, share the version you actually use (Share a prompt).

Describe the file. Do not hallucinate the answers.

Question is the coding job: themes for one open-end, or a nested scheme. Sources you may use: the instrument, the codebook from last wave if you have one, the CSV description. Pastes: column names, value labels, a sample of rows if you are allowed to paste them, and any coding rules already approved. Source policy: description and pasted rows only. No "typical NPS comments." Output shape: code list with definitions, inclusion and exclusion, and examples only from pasted text. Off-limits: invented quotes, invented percents, respondent names.

If you pasted a description without rows, the generator can propose a starter scheme from the question wording and column names. It cannot attach example quotes. Example cells stay NOT IN INPUTS until you paste real, de-identified answers.

If you pasted twenty rows, examples may quote those twenty. Do not polish spelling into a cleaner sentence and call it a quote. Do not merge two respondents into one punchy line.

No invented quotes, percents, or inter-rater scores

Theme shares, kappa, and "top mention" rankings need counts from Pastes. If you did not paste frequencies, do not print 18 of 40. Do not print a percent. A codebook is a scheme, not a results slide.

Codes: name, short definition, when to apply, when not to apply, and an example only if that example is in Pastes. Do not add Other as a dumping ground unless your rules say so. Do not add a Sentiment code unless the instrument asked for sentiment.

Hierarchy: parent and child codes only if you asked for nesting. Do not invent a three-level ontology because it looks professional. If last wave's codebook is in Sources, map new codes to old codes only where the definitions overlap. Do not force a 1:1 map.

De-identify first. Strip emails, phones, school names, and employer names from examples. If a pasted row contains a medical or HR story, add not-advice and keep the example at the code definition, not the story.

Method notes and a scheme you can dual-code

Method should say whether the scheme is inductive from pasted rows, deductive from the instrument, or a hybrid. If you pasted five rows, it is not grounded theory. It is a draft. Label it a draft.

Decision rules: mutually exclusive or multiple codes per response. If you did not specify, the generator should ask, not pick. Missing data: blank vs. "n/a" vs. refuse. Use the missing labels that are in the CSV description.

Gaps: missing row sample, missing language of responses, missing whether translators coded, missing whether the question was required, missing max length. Print those. Do not fill a sample with realistic fake answers "for illustration." Illustration is how invented quotes sneak into the slide.

Banned words: best, guaranteed, #1, invented percents. "Most respondents feel" is a results claim. Cut it from a codebook.

A worked pass: columns, no quotes yet

Say the CSV description has an open-end called Q12_other, a deadline of Friday, and an unapproved budget for extra coding. No customer quotes in the paste. The codebook can define codes from the question text and from any closed-end codes in adjacent columns. It cannot open with three testimonials. The NO_DATA list should include example quotes, theme percents, and inter-rater stats.

If a later paste includes de-identified rows, rerun and attach examples only from those rows. Do not keep the invented illustrations from the first run. The source map should change first.

Watch double-barreled answers. If a pasted row mentions price and support, the scheme needs a rule: two codes, or a primary-code rule you already approved. Do not invent a compound code PriceAndSupport unless that code is in last wave's book.

If the instrument is in another language and you did not paste a translation protocol, do not translate examples. Keep them as pasted. Mark translation as a gap.

Fill the card, then run

Paste the column description, the question text, and any de-identified sample. Leave Off-limits explicit: no minted quotes.

Question: [Question]
Sources I may use: [Sources]
Pastes / files: [Pastes]
Source policy: [Policy]
Output shape: [Shape]
Off-limits: [OffLimits]

A codebook is a contract between coders. When a run refuses to invent a quote because the CSV description had no rows, share that filled card.