Plan a customer interview round that can change one product marketing decision

Start with the decision the interviews must inform

Write the decision, its owner and its real deadline, then the learning question that connects the choice to the evidence you lack. "Which benefit leads the spring launch, for which user, based on how recent admins experienced setup" can be answered by interviews; "understand our customers better" cannot. The Product Marketing Alliance's note on planning productive customer interviews makes preparation and a consistent structure the condition for a productive conversation.

Test the wording by asking whether the opposite answer would change what you do. If the question only seeks endorsement of a choice already made, rewrite it before you recruit. Ask Claude to point out ambiguous terms and to draft the learning question, and judge the result yourself, because it works only from the facts you supplied.

Work in the Claude Project you keep for product marketing if this is a recurring job; a plain chat with the same brief attached is the fallback. Anthropic's Projects page says free accounts can create up to five Projects (checked September 15, 2026). No Research, Code or connector is needed.

The plan has these fields:

• Decision, owner, deadline, and the learning question in one sentence.
• Participant criteria per group: role, the relevant event and its timeframe, segment, exclusions.
• Recruitment source per group, and the experience the available lists do not contain.
• Invitation text, the information participants receive before agreeing, and who keeps the consent record.
• Interview guide: one open question per learning question, with probes kept separate.
• Note format, fixed before the first call.
• Review date after the first round, and the gaps to check on that date.

Recruit people whose experience can answer the question

Map each learning question to the people who lived the relevant event: recent evaluators, everyday users, administrators or former users, depending on what you need to learn. GOV.UK's service manual advises deciding who to research with based on what you need to learn.

An illustrative mapping for the setup question above, invented for this lesson:

Where the names come from limits what the round can answer. The Product Marketing Alliance's guide to customer interviews for product marketers warns that the customers your CS team trusts are likely the happiest with the product and unlikely to be a representative sample of the industry. Include contrasting experience when the learning question requires it; a study of why setup stalls needs people who stalled. A product-use question does not need lost buyers added by rule.

Paste the table without names and ask Claude which learning questions each group can answer and which important experience is missing. That surfaces omissions in what you wrote down.

Choose a first round you can conduct and review, and write the reason down. No validated number exists. Sairu, a Japanese B2B consultancy, suggests three to five companies per segment at 30 to 60 minutes each for its existing-customer interview method (in Japanese), and that is attributed guidance for that population. Track invitations, acceptances and completed interviews as three separate counts. Plan a checkpoint after the round: which groups are still absent, which answers were too generic, and whether conflicting accounts call for another group. Three similar interviews in a row do not prove saturation, and several people from one company are one company's account.

Make the invitation and consent process clear

The invitation names the sender and the purpose, why this person's experience matters, a duration you will honour, the format, and a way to choose a time. Offer a reply option beside any calendar link, and send it from someone who is not selling to the recipient. Ask Claude for a draft, then cut the sentence that explains the company.

Consent is separate from the invitation. Following your organisation's approved process, decide what participants are told before they agree: the purpose, whether the call is recorded, transcribed or processed by an AI tool, who sees the findings and in what form, how long material is kept, and how to withdraw. GOV.UK's guidance on getting users' consent for research accepts verbal consent on the recording once the participant has read the consent form you sent, and asks you to keep a record of consent. It is UK public-sector guidance, and your own requirements come from your organisation's process. The interviewer must be able to explain the arrangement in plain words.

This method uses Claude chat without connected tools, so people do the outreach, scheduling, consent and interviewing. Anthropic's connectors page says connectors let Claude take actions in connected services; that is a separately configured workflow with its own permissions, and no enabled tool replaces consent or a real participant.

Use a short guide and leave room to probe

Turn each learning question into one neutral question about a recent event. Open with role and context, then the sequence of what happened, the alternatives, the criteria, the outcome and what stayed unresolved. Keep probes such as "what happened next?" in a separate column. Nijibox's guide to user interviews (in Japanese) recommends prioritising open-ended questions; closed factual questions still have a place for dates and roles.

A good guide leaves room for an answer you did not expect. A bad one asks the participant to praise the product, or predicts buying behaviour from hypotheticals. Ask Claude to mark questions that carry an assumption or combine two jobs, then read each question aloud and time a pilot with a colleague. Keep the guide short; no source here establishes an optimal question count. For a won or lost deal, the win-loss interview guide owns the deal-specific wording and the note sheet; use the general shape here for everything else.

Decide how to preserve evidence before the first call

Fix the note format now, because analysis can only use what was recorded. Each entry carries the interview ID, segment and role, the event and its date, the question ID, a verbatim passage or a paraphrase labelled as such, your observation, your interpretation, any contradiction with another interview, the implication for the decision, and what remains unresolved. Keep the contact sheet apart from the de-identified notes you give Claude; removing names does not make a rare role or a distinctive story anonymous.

After the round, paste the de-identified notes and ask Claude to sort them into that format and to list themes with the interview IDs behind each one. Open the original note for every material theme before you use it. Report counts as "within these completed interviews" and say who was asked that question; keep dissent and non-answers beside the theme they complicate. A tidy theme without a supporting passage is a candidate interpretation. For transcripts of win-loss interviews, analysing win-loss interviews goes further.

If only friendly customers accepted, narrow what the findings can claim or recruit the missing experience; do not ask Claude to supply contrasting views. If notes cannot support exact quotes, keep paraphrases without quotation marks. Claude can rehearse question wording as a labelled pilot; nothing said in a rehearsal is customer evidence.

Review the gaps and keep the next round reusable

Write a short memo for the decision owner: the decision, what these interviews showed, what it changes, and what remains unknown, the groups you did not reach included. Feed it into the positioning brief or the launch brief it was meant to inform. If a question is still open, decide whether another group, another round, or existing public reviews can close it.

Save the approved plan, guide and note format as dated project knowledge with an owner and the remaining gaps, and check the next round against those exact files. Anthropic's Project management page says context is shared across a Project's chats through the knowledge you add, and its memory page says each Project has its own memory space. Memory may carry context between chats; a dated file with an owner is the research artefact a reviewer can open.

One prompt earns its place, as a scope check before you invite anyone:

Using this decision and these participant criteria, list the evidence each group could provide, the important experience missing from the sample, and any question that assumes our preferred answer. Mark unknowns. Do not invent participants or findings.

When asynchronous feedback can help

Nothing above needs software. The step that lapses after the guide is approved is operational follow-through: outreach, scheduling, consent records, the interviews and the dated write-ups. The DIY remedy is one owner, a contact and status sheet, the approved invitation and consent process, and analysis files with dates.

Steve Win/Loss fits a narrower job. You create a win, loss or product-feedback survey, review and edit the generated questions, then send the survey link to recipients you choose manually; for closed-deal win/loss, a synced CRM deal can trigger the invitation after a delay you configure. The buyer answers in an asynchronous AI conversation that asks follow-up questions to clarify a response, and you generate a summary when you want one. The tradeoffs, as of September 2026: it supplies no decision, sampling plan, recruitment relationship, booking or live interviewer and treat the product-feedback mode as something to test in your own account before relying on it; and collected answers are records from the people who responded, with no representative sample behind them. If asynchronous collection with reviewed questions is the step you cannot sustain, see Steve Win/Loss.