Synthesize six user-interview excerpts into a themes doc where every theme carries verbatim, attributed quotes — then audit the AI's draft line by line. One contradiction is buried in the transcripts, and the AI will almost certainly assert one theme no transcript supports. Catching both is the graded outcome.
Rohan is a PM at a Bengaluru GCC who just took global ownership of the onboarding surface of Helixware, a fictional US-headquartered HR platform used by mid-market companies to onboard new employees. Discovery week produced six interviews across three customer roles. The VP wants a themes readout tomorrow. Rohan's plan — and yours — is to let AI do the first synthesis pass, then run a citation audit so ruthless that nothing reaches the readout unless a transcript line can be pointed to for it.
Six excerpts, condensed to their load-bearing lines. Paste-ready.
P1 | Anika, HR admin, 400-person logistics co.
"Onboarding one new hire takes me about 40 minutes of form-filling across
three screens. When we hired 30 drivers in one week I nearly cried."
"I keep a personal Excel checklist because I don't trust the system to
tell me what's pending."
P2 | Marcus, HR admin, 250-person agency
"The checklist reminder emails go to the hiring manager only — the new
hire never gets them. That's by design, I checked with support."
"Half my managers ignore those emails. I chase them on Slack."
P3 | Devon, hiring manager, 600-person retailer
"I get an email storm every time I hire someone. I want one weekly digest,
not eleven notifications."
"IT equipment requests live in a different tool. I forget every time."
P4 | Sara, new hire (joined 3 weeks ago), fintech
"Day one I had no laptop and no idea who to ask. My manager was traveling."
"The tasks I could see were HR forms. Nothing about my team or my work."
P5 | Lena, HR admin, 350-person healthcare co.
"Bulk import exists but it errors out on our custom fields, so we gave up
and type each hire in manually. Forty hires a quarter, typed by hand."
"What I actually want is for the offer-letter data to just flow in."
P6 | Tomas, new hire (joined last month), logistics co.
"I got seven checklist emails from Helixware in my first week. Seven.
I stopped reading them by Wednesday."
"Nobody told me my badge and parking had to be requested separately."
Read the transcripts yourself first — ten minutes, no AI. Note, in your own words, the three or four patterns you see and any line that surprises you. This is your anchor: if the AI's synthesis later contains a theme you never saw evidence for, you want the instinct to fire. Do not skip this; the whole lab depends on you having a pre-AI opinion.
Get the AI's first-pass synthesis. Paste all six excerpts with this prompt: "You are a product researcher. Synthesize these six interview excerpts into exactly five themes. For each theme: a one-line title, a two-sentence summary, and the supporting verbatim quotes with the participant ID. Do not include any theme without quote support." Asking for exactly five is deliberate — hold that thought.
Build the citation-audit sheet. Make a table with columns: Theme | Quote as cited | Participant | Found verbatim? (Y/N) | Actually supports theme? (Y/N) | Notes. Copy every quote the AI cited into it. This sheet — not the themes doc — is the artifact this lab grades.
Audit every citation against the source. For each row, find the quote in the transcript. Check two things separately: the quote exists word for word (paraphrase counts as N — AIs routinely tighten quotes into things nobody said), and the quote actually supports the theme it is filed under. A real quote pressed into service for an unrelated theme is the subtler failure. Mark every N and write why.
Hunt the unsupported theme. You asked for five themes; count how many survive the audit with at least two genuinely supporting quotes from at least two participants. Any theme resting on zero verbatim quotes, one stretched quote, or a plausible-sounding generalization no one said — flag it as UNSUPPORTED in the sheet. If all five somehow survive, re-read the weakest one with maximum suspicion before you clear it.
Hunt the contradiction. Prompt: "List every factual claim in these transcripts that another transcript contradicts. Quote both sides." Then verify its answer against the source yourself — and check the specific pair the AI is most likely to smooth over: what do P2 and P6 each say about who receives checklist emails? Write the contradiction into the sheet as its own row, with both quotes.
Rebuild the themes doc honestly. Ask the AI to regenerate the synthesis using only audited quotes, with the unsupported theme removed and the contradiction listed under a heading called "Conflicting evidence — needs follow-up" rather than resolved by guesswork. Add one line per theme in your own words on what you would do next (follow-up question, data pull, or design probe).
Write the two-line audit summary. At the top of the sheet: how many citations audited, how many failed and why, one contradiction logged. This is the line you would say out loud in the readout before presenting themes — it is what makes the synthesis trustworthy.
Ask for one more theme than you believe the data holds — the model will pad rather than push back, and watching where it pads teaches you exactly what unsupported synthesis looks like.
A complete lab has all of the following:
AI synthesis is a draft of an argument, not evidence. The citation audit — verbatim check, relevance check, contradiction hunt — is what turns forty minutes of AI output into a readout you can defend line by line.