Skip to content
AITroveRead. Build. Understand.
Reading list
Make this comfortable

Interview prompts: write a codebook with inclusion rules

Last updated: 5 Oct 202611 min read
tutorial
AdvancedBy AITrove Editorial

A code label is too vague to support consistent analysis. Give each code an operational definition, an inclusion test, an exclusion test, a representative excerpt, and a near miss. Let a segment hold more than one code when independent claims support them. Keep a version number for the codebook and distinguish new inductive candidates from approved codes. Ask the model for the segment ID, candidate code, evidence span, and reason; a researcher approves additions and ambiguous assignments. Do not interpret a broad word such as 'late' as a missed pickup if the speaker only describes a slow page.

Operational case

The Parcel Window team defines WINDOW_CLARITY for uncertainty about the pickup interval shown to the customer. P-07 says the app displayed 'morning' without an end time, so the segment qualifies. P-02 says the confirmation page loaded slowly but understood the interval; that is a near miss. An early model draft tags both as WINDOW_CLARITY. The researcher rejects the second assignment and adds a separate PAGE_LATENCY candidate only after checking the rest of the interviews.

Output
Code WINDOW_CLARITY v2
Include: uncertain displayed pickup interval
Exclude: page delay with understood interval
P07-S11 -> include, missing end time
P02-S06 -> exclude; PAGE_LATENCY candidate

Performance and review cost

Applying K candidate codes to N segments can require O(NK) comparisons in a direct pass. A compact codebook reduces repeated interpretation but may miss a genuinely new pattern, so reserve a candidate-code lane and review it periodically. Version changes require recoding affected segments; the cost is explicit and preferable to mixing incompatible definitions in a single count.

Common Mistakes

  • Do not use a label without an inclusion and exclusion test.
  • Do not force every new observation into an existing code.
  • Do not compare counts from different codebook versions without recoding.

Connected lessons

prompt engineering
qualitative research
Storage details