Analyze win-loss interviews with Claude and trace every theme back to the buyer's words

Define the interviews you are analyzing

Make a manifest before you read anything. One row per interview: filename, interview ID, deal ID, date, won or lost, and why it is included. Mark any record that is missing or unreadable. Interviews, buyers and deals are different units, so if you spoke to two people about one deal, or to the same buyer twice, decide which unit you will count and say so in every finding. An interview added later goes in as a new version, with the counts recalculated.

The manifest also defines the cohort. "Lost mid-market deals this quarter" is a denominator, and so is the whole set. Either is legitimate when the finding names the one it uses.

Use permission-cleared transcripts or survey exports with speaker labels and stable IDs, and remove identifiers the decision does not need. If you have no interviews yet, start with running win-loss without a research team and writing the interview guide.

For a study you will repeat, keep the manifest, the label definitions and the checked outputs in a Claude Project; the project setup lesson covers that once. Anthropic's Project management page says context is not shared across a Project's chats unless it is added to the knowledge base, so save outputs as files. For a single-pass study, you can attempt the work in an ordinary chat with approved files attached. Corpus size and coverage still require the manifest and per-interview checks below.

Extract buyer passages before naming themes

Ask for the decision-relevant buyer passages from one interview, each with the exact text, the interview ID, a locator such as a paragraph number or timestamp, and a short provisional label. Keep enough surrounding words to preserve who is speaking and any qualification they added. Ask for the quotation and Claude's interpretation as separate fields, with ambiguity marked rather than resolved.

Clozd's guidance on tracking themes asks for the same thing from the sales side: each time an interview is tagged with a theme, clip a quote in the buyer's own words. A label with no passage under it is unsubstantiated in that output, which is a different thing from invented. Go back to the interview before you keep or drop it.

One prompt is worth keeping, because it fixes the unit and separates quotation from inference. It has not been benchmarked; treat it as a starting wording.

Analyze interview [ID] for [decision question]. Extract relevant buyer passages verbatim with source locations and provisional labels. Keep qualifications and conflicting statements. Separate each quotation from your interpretation; mark ambiguity instead of filling it in. Use only this supplied interview. Do not treat an omitted passage as evidence of absence.

Check every passage you intend to use against the original. Reading the beginning, middle and end is an early spot check, not proof that the intervening material was processed. Reconcile each interview's completion entry against the manifest. Checking three quotations early can catch a paraphrase presented as a quote; it does not verify the quotes you use later, so keep checking as you go.

Group by decision meaning and keep the exceptions

Calibrate on a small initial subset before expanding. Decide whether two passages describe the same decision factor: "too expensive" can mean the price was high next to an alternative, the value was unclear, or the budget was gone. Split a label only when the passages support the distinction; otherwise mark it unresolved. Write each label with an inclusion and an exclusion note, and keep evidence that fits no label.

Starting from the passages, with labels drawn from the buyers' wording, is this lesson's recommendation for a first pass. A declared codebook is also legitimate for a targeted question, as long as it can be revised and each assigned label is checked against the passage it sits on. Reject a theme when the passage does not support the label, never because you supplied the category.

Goodpatch's research team reported a GPT-4 test on remote-working interviews in which a prompt reflecting their researchers' practice produced needs-based labels with the speech fragments retained, where a generic prompt gave word-level groupings with summarized descriptions; their conclusion was to use the output as a たたき台, a first draft the researcher refines (Goodpatch, in Japanese). That was one team's experiment with a different model. What transfers is the discipline: keep the fragments attached and check every grouping.

Apply the settled definitions across the rest of the manifest, one interview at a time, with a completion log. A long interview split into chunks is still one interview. When a definition changes, revisit the rows already coded, and keep raw quotes, summaries and interpretations in separate columns.

Count interviews within each theme

For each theme, give every interview in the cohort one of five values, supports, contradicts, mixed, not addressed or unclear, with the passage that justifies it. "Not addressed" means you inspected that interview and the topic did not come up; it is never inferred from a passage the model failed to return. Then count unique IDs per value.

An illustrative example, invented for this lesson: five interviews about a pricing decision.

The finding reads: in these five interviews, two support the price theme, one contradicts it, one is mixed and one did not address it. I04 is not silently counted as a supporter. A separate implementation theme might include I01 and I04, and adding the two themes' counts would count those interviews twice, so totals across themes need not sum to N. A buyer who names a different concern has not rejected this one.

Small sets make percentages misleading. With nine interviews, one interview moves a theme by 100 divided by 9, about 11.11 percentage points, so "44% of buyers" claims a precision the set does not have. Write the count, the denominator and the exceptions in one sentence, so the theme cannot travel without them. Dovetail's win-loss guide recommends recording each interview's themes in a consistent format so patterns can be checked across the full set. The denominator and precision rule here is this lesson's editorial method.

Check the evidence before sharing a finding

• Every quoted passage was read in its source at the location the sheet gives; paraphrases sit in the interpretation column without quotation marks.
• Every interview on the manifest has a completion entry, with its beginning, middle and end inspected rather than inferred from the output.
• Each theme label fits the passages under it, whether it came from the passages or from a declared codebook, and the codebook version is recorded.
• Each finding names its cohort and denominator, with the supports, contradicts, mixed, not-addressed and unclear counts for that cohort, in unique interviews, buyers or deals as declared.
• Unresolved ambiguities are listed as open questions rather than assigned to the nearest theme.

Save the manifest, the versioned codebook, the checked quote sheet and the unresolved cases, so the next study revises the codebook instead of rebuilding it.

Know what the records cannot establish

The transcript records what the buyer said. It does not establish an unstated motive, and a coherent explanation from Claude does not make one true. Cotion's guide to loss analysis separates AI output into facts, trends and hypotheses (in Japanese); keep the third layer labeled as hypothesis until a buyer confirms it.

Nor can the set establish prevalence beyond itself. A finding says "in these interviews"; more interviews do not correct a biased selection, and a larger context window does not change that. A theme that appears once but describes a consequential failure stays on the sheet as a follow-up question. Frequent themes are candidates for messaging work, which the win-loss to messaging lesson takes up, and frequency alone does not decide what matters. When the CRM records a different reason from the buyer, compare the two on the same deal rather than overwriting either.

When structured survey collection is the next job

Keeping the buyer's side arriving, period after period, can be harder than producing one theme sheet. The DIY remedy is a reusable interview guide, fixed metadata fields, an owned codebook, per-interview review and a stated cadence.

Steve Win/Loss covers a narrower job. It collects structured buyer answers through an AI chat that can ask follow-up questions, and an operator can generate a summary over those stored answers when they choose to, covering themes, loss drivers and competitor insights. The summary is built from the responses available when you click Generate Summary; check which responses you have approved. Treat it as a survey record rather than a verbatim transcript quotation. The Win/Loss survey and report workflow described here does not import or code an existing transcript corpus, and the interpretation, quote checks and sampling decisions in this lesson stay with you. If collection is the part you cannot sustain, see Steve Win/Loss.