Full SKILL.md
Primary Research Synthesis
Designs interview programmes and synthesises real interviews, expert calls and observations into a codebook, themes with evidence counts, insight statements, verbatim quote bank and espoused-versus-in-use gaps (Grounded Theory coding; Argyris and Schön). Use whenever transcripts, call notes or observation notes must become findings with evidence behind them, or the user says "synthesise these interviews", "build an interview guide", "code these transcripts", "themes from the expert calls", "what do people actually do versus what they say", "is fifteen interviews enough", "pull the quotes for the deck". Not for documents and board packs (document-synthesis) or external desk research (boardroom-storm-research).
Runs collection and analysis together so the sample adapts to the data, and delivers findings that carry their own counts, sources and counter-evidence. Method sources are in references/frameworks.md; the guide structure, question bank, Model I probes and observation protocol are in references/interview-guide.md.
When not to use
- Documents, data rooms, prior decks:
document-synthesis. - External sources and simulated perspectives rather than real interviews:
boardroom-storm-research. - Hypotheses need structuring before fieldwork:
hypothesis-stress-testing. - Findings need to become a storyline:
executive-communication.
Inputs
Ask for, or infer and label: the one-sentence question and the decision it informs; hypotheses in play with what would disconfirm each; who has been or can be interviewed (segments, counts, seniority, sites; fixed vs open); non-interview data available (logs, records, observation access); constraints (timeline, anonymity, recording, client sensitivities); raw material if fieldwork is done.
Ask one round of questions if the question or the sample is unknown. Otherwise state segment sizes, stop rule and attribution level as assumptions and proceed.
Workflow
-
Design the programme (
references/templates.md§1). Segments: everyone inside the mechanism (each party in the decision chain, system owners, at least one counterparty). Name at least one deviant case on purpose. Plan a non-interview source per expected theme. Write the stop rule now: per segment, three consecutive interviews adding no new codes and no new properties to the top three themes, minimum six before claiming saturation. Reserve 30–40% of slots for allocation after the first coding pass. Test: if the answer turns out to be about incentives, who in this sample would know? -
Interview guide. Funnel: role and context; the topic in their words; two or three recent incidents in detail; conditions and variation; unscripted follow-up. Pair every policy question (E, espoused) with an incident question on the same topic (U, in use). Add Model I probes: who gets asked when it goes wrong, worst personal outcome, what you are measured on. Close with "who would disagree". Review test (
references/interview-guide.md§5): rewrite anything yes/no, leading, hypothetical, or outside the segment's line of sight. Revise after interview three; keep a version log. Done when every topic has an E and a U question and a quarter of the time is unscripted. -
Fieldwork with coding from day two. Two interviewers where possible; record with consent; transcribe within 24 hours. Ten-minute surprise note after each interview: unexpected, contradicted, add to guide. Open coding starts after interview two. No emerging-conclusion briefings to the client mid-fieldwork.
-
Open coding. Line by line on the first four to six interviews. Multiple codes per incident; use the interviewee's words where sharper; tag E or U. Codebook entry per code (name, definition, include, exclude, example); no exclude column means the code is too broad. Expect 40–80 codes by interview six. When two incidents in one code conflict, write a memo (
references/templates.md§5): the conflict, suspected relationship, what would disconfirm it, who to sample next. Done when a second coder could apply any code without asking. -
Constant comparison. Compare incidents against the properties of their category: what varies (speed, frequency, who, condition). Merge codes that are properties of one thing; drop codes that connect to nothing; log every merge so counts stay reproducible. 20–30 codes from 60 is normal. Segment check: which segments produced each code, and where a code is absent from a segment that should have seen it.
-
Saturation and reserved slots (
references/templates.md§6). Log per interview: segment, new codes, cumulative, and the judgement column "new properties on top themes?". Code saturation arrives before meaning saturation (Guest et al. 2006; Hennink et al. 2017 as calibration, not rules). An interview that adds no code but changes the core category counts as new. Allocate reserved slots by the memos: deviant case, the segment that can confirm or refute the core, the counterparty. A segment under six reports "consistent across the n interviewed", not "saturated". -
Axial coding. For the top four to six categories, the paradigm table: phenomenon, causal conditions, context, intervening conditions, strategies, consequences, each row referenced to interviews, observations and data. Leave a row blank rather than fill it by inference. State each theme as a claim ("the approval matrix routes discounts upward without changing the decision rule"), not a topic.
-
Selective coding and negative cases. Candidate core: appears most, connects most, explains variation. Test against every major theme and every negative case; a candidate that cannot explain the deviant site is not the core. Storyline of three to five sentences relating each theme to the core. Negative cases get their own section; they are usually where the lever is.
-
Espoused vs in-use (
references/templates.md§10). Per topic: one sentence each with evidence. Classify the gap: unaware; known and tolerated; known and concealed; policy wrong or ambiguous. Name the governing variable sustaining it ("win, do not lose" first in commercial settings). Weight on three separate dials: frequency (n by segment), intensity (unprompted, emphasised, returned to), position (placed to observe what they describe: front line for in-use, leaders for espoused, customers for external posture; seniority is not a weight on truth). Triangulation overrides all three: interviews plus observation plus record is a finding; interviews alone is a reported view, and the pack says which. -
Insights, quotes, pack. Each insight: one sentence on mechanism; count (n of N by segment); observations and data; intensity; counter-evidence; strength separately for existence, magnitude, prevalence; what would overturn it; implication; upstream hypothesis supported / not supported / not testable. Quote discipline: verbatim (fillers removed, ellipsis for cuts, nothing else changed; otherwise labelled paraphrase); attributed by role and site only, one level up for segments under three; never composite; traceable to interview and line range; representative first, minority quotes labelled as minority; clearance columns honoured.
Output
Governing insight in the first 150 words. Then, each once (references/templates.md §12):
- Governing insight and storyline
- Insight statements with counts, sources, counter-evidence, strength, implication
- Theme hierarchy with evidence counts by segment
- Espoused / in-use gap table with classification and governing variable
- Negative cases
- Hypothesis verdicts
- Method appendix (design sheet, saturation log, codebook summary, merge log)
- Quote bank with attribution and clearance columns
Working files (transcripts, full codebook, memos) stay behind the pack. No method vocabulary, framework authors, skill names or the worked example's client in the deliverable. Full example: references/worked-example.md.
Quality bar
Run silently; fix, do not report.
- Every insight has a count by segment and a counter-evidence line.
- Every finding labelled as such has a non-interview source; reported views are labelled reported.
- Every quote is verbatim, single-speaker, traceable and attributed at the agreed level.
- The stop rule was written before fieldwork and the log shows it met or which segments fall short.
- Each theme is a claim with a condition the client controls.
- The core explains the deviant case.
- Position is applied per claim type, not by seniority.
- Each hypothesis has a verdict.
- Merge log makes every count reproducible.
- No leaked method vocabulary.
Anti-patterns
- Coding after the last interview instead of the second.
- Themes as topics; codes without an exclude column.
- Counting codes only; stopping when new codes stop but meanings still move.
- A director's account of rep behaviour treated as evidence of rep behaviour.
- Composite, cleaned or misattributed quotes.
- Vivid minority quote presented as representative.
- Negative cases dropped because they complicate the story.
Hand-offs
References
Frameworks: Grounded Theory and Theories of Action
3,893 wreferences/frameworks.md14 sub-sectionsInterview guide design and question bank
1,465 wreferences/interview-guide.md7 sub-sectionsTemplates and checklists
1,573 wreferences/templates.md15 sub-sectionsWorked example: Norvane Industrial Supply, margin leakage interview programme
4,548 wreferences/worked-example.md15 sub-sections