/voice

Thematic Analysis Reference (Braun & Clarke)

Purpose: Inductive coding of open-ended feedback (survey free-text, interview transcripts, review bodies) using Braun & Clarke's six-phase reflexive thematic analysis (2006, refined 2019). Surfaces latent themes, tracks saturation, and converts unstructured text into a defensible, citable thematic map — the gold standard for qualitative feedback synthesis.

Scope Boundary

• voice thematic: 6-phase Braun & Clarke coding, codebook governance, saturation tracking, inter-coder agreement (Cohen's κ / Krippendorff's α), and thematic map output. Inductive (data-driven) by default; deductive (framework-driven) on request.
• voice classify (elsewhere): pre-defined category tagging for high-volume routing. Classify uses fixed buckets; thematic discovers them. Once thematic stabilizes a codebook, hand it off to classify for ongoing tagging.
• voice sentiment (elsewhere): polarity / emotion detection per response. Orthogonal to themes — a single theme can carry positive and negative sentiment.
• field (elsewhere): qualitative interview design and sampling. Field owns study design and recruitment; thematic owns the coding methodology applied to the resulting transcripts.
• echo (elsewhere): persona-driven UX walkthrough. Thematic findings can validate or challenge persona assumptions; echo cannot substitute for grounded coding.
• insight (elsewhere): synthesis-to-recommendation. Thematic produces themes; insight turns themes into prioritized actions.

Workflow (Braun & Clarke 6 Phases)

PHASE 1   →  Familiarization — read all data twice before coding; capture initial impressions
          →  do not skip; analytic depth depends on immersion

PHASE 2   →  Initial coding — line-by-line tags on raw data; descriptive and short
          →  one segment can carry multiple codes; do not collapse early

PHASE 3   →  Theme search — cluster codes into candidate themes; build initial thematic map
          →  themes ≠ codes; themes capture meaning, codes capture surface

PHASE 4   →  Theme review — collapse, split, or drop themes against the full dataset
          →  every theme must have multiple supporting codes from multiple participants

PHASE 5   →  Theme definition — name and define each theme in one sentence; identify essence
          →  if you cannot define it in one sentence, it is two themes

PHASE 6   →  Report — narrative weaving themes with verbatim quotes; cite participant IDs
          →  themes answer the research question, not just summarize the data

Inductive vs Deductive Coding

State the mode in the report header. Hybrid is the default for product feedback — apply a deductive frame (e.g., journey stage) but allow inductive sub-themes inside each frame cell.

Theme Saturation Tracking

A theme is "saturated" when additional data adds no new codes within it. Track explicitly:

For text feedback at scale, plot a saturation curve (cumulative unique codes vs. responses coded). Stop coding when the curve flattens for two consecutive batches.

Sample Size Guidance

Below the lower bound, treat output as exploratory only. Above the upper bound, returns diminish — cohort-split instead.

2026 LLM-Assisted Coding Stack

By 2026 the discipline shifted from "code by hand" to LLM-first / human-validated coding for any dataset of 200+ responses. Practical tooling:

Quality bar from published benchmarks (2026): ~90-95% theme-detection accuracy + sentiment on English B2B feedback when the LLM is given a team-curated codebook. Vendor-default categories rarely match how the business already discusses its product — always start from your phase 4 locked codebook, not the tool's defaults.

Human-in-the-Loop Discipline (mandatory)

• Codebook ownership stays human. The LLM proposes; the analyst accepts / merges / splits / rejects each candidate code before the codebook locks.
• Spot-check Cohen's κ on 5-10% of LLM-assigned codes before publishing themes. Below κ = 0.6 agreement vs the analyst, the codebook needs another round.
• Negative-case search remains a human task. LLMs over-fit to dominant patterns; the "data that contradicts the theme" rule from the Saturation Tracking section is the analyst's job, not the model's.
• Cite the LLM transcript ID alongside participant ID for every quoted segment — the audit trail mirrors what Mend / Beacon require for AI-assisted artefacts.

Codebook Discipline

• One code = one analytic idea. Prefer "feels lost in onboarding" over "onboarding bad."
• Code names are short (≤6 words), descriptive, and analyst-neutral (no judgmental words).
• Every code carries: definition (≤2 sentences), 2-3 verbatim examples, exclusion rule.
• New codes during phase 2/3 trigger codebook revision; revisions force re-coding of prior data.
• Lock the codebook only at phase 4 — phase 5/6 changes require re-opening phase 4.

Inter-Coder Agreement

For multi-coder studies, measure agreement before reporting:

Below the "acceptable" threshold, return to phase 5 to clarify code definitions, then re-code. Do NOT publish thematic results without an agreement statistic when ≥2 coders contributed.

LLM-Assisted Coding (Augment, Don't Replace)

LLMs accelerate phase 2 (initial coding) but degrade phase 3-5 (theme construction) without human oversight (see Human-in-the-Loop Discipline above for the κ threshold and codebook-ownership rules).

• Use LLMs to generate first-pass codes on raw segments at temperature 0; require structured JSON output.
• Document LLM model, version, prompt, and seed in the report; LLM-coded studies are not reproducible without this.

Anti-Patterns

• Skipping familiarization — coding cold produces shallow, surface-level codes that miss latent meaning.
• Treating frequent codes as themes — frequency ≠ significance; a single rich quote can anchor a theme that a frequency count would bury.
• Naming themes after categories ("Onboarding") instead of meaning ("Onboarding feels like a test you can fail") — category names are descriptive, not analytic.
• One-coder studies with no audit trail — irreproducible; demand at least a second-coder spot-check on 10% of data.
• Forcing all data into themes — leaving 5-15% as "outlier / unthemed" is honest; forcing 100% theme coverage manufactures themes.
• Ignoring negative cases — themes without disconfirming-evidence search are confirmation bias dressed up.
• Reporting themes without verbatim quotes — the evidence trail collapses; readers cannot verify.
• Pooling Detractor and Promoter NPS comments into one thematic pass — scoring asymmetry produces fundamentally different themes; code separately, then compare.
• Using deductive coding without naming the framework — readers cannot tell whether a theme was discovered or imposed.

Handoff

• To voice classify: stable codebook (≥2 saturation batches) → operationalize as classifier categories for ongoing tagging at scale.
• To voice insight: theme-level findings with verbatim evidence → translation into prioritized actions and owner recommendations.
• To Field: ambiguous themes or persona divergence → escalate for moderated interviews to disambiguate.
• To Spark: themes describing unmet jobs or workarounds → feature proposals.
• To Echo: themes pointing to persona-specific friction → cognitive walkthrough validation.
• To Growth: themes correlating with churn cohort feedback → retention-program inputs.
• To Magi: cross-segment recurring themes → strategic narrative inputs.

References: Braun & Clarke (2006), "Using thematic analysis in psychology"; Braun & Clarke (2019), "Reflexive thematic analysis"; Guest, Bunce & Johnson (2006) on saturation thresholds.