What feedback files do I need, and how should I prepare them?

Gather the raw artifacts — a support export CSV, sales call notes, interview transcripts — not your own summary of them. Before anything goes into a prompt, replace every customer name, email address, company name, and account ID with a [customer-n] label. De-identifying is not optional; this is real customer PII.

How does this handle one loud customer who sends many tickets?

The playbook explicitly separates message count from distinct-customer count. Five tickets from one frustrated account is noise, not a pattern; Claude is asked to flag themes that look large only because one or two customers repeated them, and to weight every theme by the number of distinct people, not the number of messages.

What if a theme Claude finds has no real quotes behind it?

Drop it. A theme with no traceable quotes means Claude over-generalised past the data — it's not a finding. Insisting on real verbatim quotes for every kept theme is the discipline that makes the synthesis defensible; a roadmap built on a hallucinated pattern is worse than no synthesis at all.

Can I run this on a small feedback pile — say, fewer than 10 customers?

Yes, but be honest about what the numbers can support. With a small set, "3 of 8 customers" is already a meaningful signal, but name the limitation in the output. The playbook's traceability requirement — real quotes per theme — is even more important with small samples, where one vivid story is easiest to mistake for a trend.

Synthesize feedback into a roadmap signal

when to reach for this

Customer feedback arrives as exhaust — support tickets, sales call notes, interview transcripts, a Slack channel of one-liners — and the loudest request usually wins by being loud, not by being right. This system turns that pile into a roadmap signal: the themes that genuinely recur, an honest read on what's a real pattern versus one customer who emailed five times, and a short list of bets worth making. You do it all by dropping the files into the chat in Claude Desktop — no terminal. The discipline that makes it trustworthy is traceability — every theme has to point back to the actual quotes that produced it, so a roadmap built on it can survive the question "says who?"

gather this first

The raw feedback as files — support-export.csv, sales-notes.md, interview transcripts — dropped straight into the Claude Desktop chat. Real artifacts, not your summary of them, so the patterns are Claude's to find, not yours to confirm.
Roughly how many distinct customers each source represents, so "recurring" can be weighed by people, not by message count.
What you'd do something about — "we can change the product and the onboarding, not pricing this quarter" — so the synthesis points at actionable themes.

the workflow

De-identify, then have Claude read it back as themes — not solutions

Strip names and contact details first, then ask only for the patterns. Jumping to "what should we build" before the themes are clean lets one vivid story masquerade as a trend.

you ask

Here's customer feedback from support-export.csv, sales-notes.md, and two interview transcripts — I've replaced names and emails with [customer-1], [customer-2], etc. Read all of it and give me the 5-7 recurring themes, each in one line. Just the themes and how often each shows up — no solutions yet.

what you get back A ranked theme list with rough counts — "onboarding confusion (mentioned by ~9 of 14 customers); wants a save/bookmark feature (~6); export to Excel (~4); slow search (~3)." Patterns surfaced from the raw text, not your priors confirmed.

De-identifying first isn't optional — this is real customer data. Replace names/emails/account ids with [customer-n] before any of it goes into a prompt.
Separate a real pattern from a loud one-off

This is the judgment that earns the playbook. Five messages from one frustrated account is noise dressed as signal; the synthesis has to weight by distinct people, not volume.

you ask

For each theme, tell me whether it's a real pattern or a loud minority: how many DISTINCT customers raised it (not how many messages), how strongly each felt, and whether it skews to one segment. Flag any theme that looks big only because one or two customers repeated it a lot.

what you get back A signal-vs-noise read — "save feature: real — 6 distinct customers across segments. Custom theming: loud one-off — 1 enterprise account, mentioned in 4 tickets. Slow search: real but small — 3 customers, all power users." The count of people is the truth, not the count of words.
Trace each theme back to the quotes

A theme you can't trace to real words is one you can't defend on a roadmap. Pulling the receipts also catches Claude over-generalizing from a single colourful line.

you ask

For the top 4 themes, pull 2-3 representative verbatim quotes each (keep the [customer-n] labels), with which source they came from. If a theme doesn't have real quotes behind it, say so and drop it.

what you get back Each kept theme backed by real quotes — "onboarding confusion: '[customer-3], support: I couldn't tell how to get started after signing up.'" Anything Claude can't substantiate gets flagged and cut. This is your audit trail.

If a theme has no quotes, it was Claude pattern-matching past the data — drop it. The receipts are what make the roadmap defensible.
Turn the signal into a few bets

End with a short, opinionated shortlist, not all seven themes. A roadmap signal is a recommendation about where to spend, and every bet stays tethered to the evidence behind it.

you ask

Recommend 3 bets for the next quarter, ranked. For each: the theme it addresses, how many distinct customers it'd serve, rough effort (small/medium/large), and the quotes that justify it. Then name one theme that's tempting but I should NOT chase yet, and why.

what you get back A ranked shortlist with receipts — "1. Fix onboarding (9 customers, medium effort, quotes attached); 2. Add save feature (6 customers, small); 3. Export to Excel (4 customers, small). Don't chase yet: custom theming — one loud account, large effort." A signal you can take to the roadmap and defend line by line.

make it your own

Feed the spec: take the top bet straight into Pressure-test an idea into a v1 spec, and the customer quotes become the user stories — the demand and the spec stay tethered.
Quarterly cadence: keep a feedback/ folder and re-run each quarter, asking Claude to diff against last quarter's themes so you can see what's rising, fading, or newly fixed.
Always-on intake (Power Track): for steady inflow, a scheduled agent (see the Capabilities tab) can pre-cluster new tickets weekly so the quarterly synthesis starts from a tidy pile, not a raw dump. Scheduled agents are an opt-in Power-Track step — on Desktop you can just re-run the synthesis by hand each quarter.

watch out for

Claude finds patterns; you weigh them. "6 customers want X" is an input to a roadmap call, not the call — strategy, effort, and fit are yours to own. Don't let a tidy theme list make the decision for you.
Customer feedback is PII. De-identify before it enters a prompt — strip names, emails, company names, and account ids, and never paste a raw support export with contact details into anything you wouldn't show the whole customer list.
A theme with no quotes behind it is Claude over-generalizing, not a finding. Insist on traceability and cut anything that can't show its receipts — a roadmap built on hallucinated consensus is worse than no synthesis at all.

you'll end up with A defensible roadmap signal — recurring themes weighted by distinct customers, a real-pattern-vs-loud-one-off read, and 3 ranked bets each traceable to the exact quotes that justify it — so the next quarter is set by evidence, not by who emailed loudest.

Questions people ask

What feedback files do I need, and how should I prepare them?: Gather the raw artifacts — a support export CSV, sales call notes, interview transcripts — not your own summary of them. Before anything goes into a prompt, replace every customer name, email address, company name, and account ID with a `[customer-n]` label. De-identifying is not optional; this is real customer PII.
How does this handle one loud customer who sends many tickets?: The playbook explicitly separates message count from distinct-customer count. Five tickets from one frustrated account is noise, not a pattern; Claude is asked to flag themes that look large only because one or two customers repeated them, and to weight every theme by the number of distinct people, not the number of messages.
What if a theme Claude finds has no real quotes behind it?: Drop it. A theme with no traceable quotes means Claude over-generalised past the data — it's not a finding. Insisting on real verbatim quotes for every kept theme is the discipline that makes the synthesis defensible; a roadmap built on a hallucinated pattern is worse than no synthesis at all.
Can I run this on a small feedback pile — say, fewer than 10 customers?: Yes, but be honest about what the numbers can support. With a small set, "3 of 8 customers" is already a meaningful signal, but name the limitation in the output. The playbook's traceability requirement — real quotes per theme — is even more important with small samples, where one vivid story is easiest to mistake for a trend.