The part of customer synthesis I still do by hand

Recording, transcript, the prompt I run, and the five minutes at the end that a model cannot do for me. Yet.

My week has a shape. Monday and Tuesday, I am on calls. Content-developer syncs on AZ-400 labs. A partner call with a training provider who wants their instructors ready for another course before the exam refresh. A learner interview panel where five people who took the cert last quarter tell me, on camera, what they hated. By Friday I owe a one-page synthesis to my PMM lead and the skilling director. Signal in, decisions out.

For years, the middle of that week was the worst part of the job. Not the calls. The synthesis. Twenty pages of notes, four transcripts, my own memory decaying by the hour, and a blank doc titled “what did we hear this week.”

That middle part is now mostly a model. But not entirely, and the “not entirely” is the whole point of this piece.

The workflow, end to end

Here is what I actually do. No theory. The order matters.

1. Record everything, one place. Every call gets the Teams recording plus the transcript export. Partner calls, instructor calls, learner panels, internal stakeholder calls with the content-developer team. One folder per week, named by the ISO date. If it is not recorded, for me it did not happen, because I will not trust my memory of a 55 minute conversation on Thursday when I sit down to write on Friday.

2. Clean the transcript, once. Before any model touches it, I open the transcript and do two things by hand. I fix the speaker labels (Teams still confuses two people with similar names), and I delete the first 90 seconds of small talk. Both take about a minute. Both matter, because a model reading a transcript that says “Speaker 2” 40 times will confuse the partner and the learner, and a model reading two minutes of weather chat will pick up “delay” and “frustrated” from a sentence about traffic and quote it back at me as a customer signal.

3. Run the synthesis prompt, per call. Not per week. Per call. Batching four calls into one prompt sounds efficient and produces mush. One call at a time, same prompt, structured output. This is the workhorse:

You are helping a Senior PM synthesize a single stakeholder call.
The PM owns labs and interactivity experiences strategy for Microsoft Global Skilling. The person on the call is a [role: content
developer / partner / learner / instructor]. Their relationship to
the product is [one sentence].

Read the transcript below and return exactly four sections.

1. Verbatim quotes worth keeping (max 5). Speaker attribution and
   a rough timestamp if visible. No paraphrasing. If nothing meets
   the bar, return fewer than 5. Do not pad.

2. Decisions or asks the person made of me, explicit or implied.
   For each, mark E (explicit, they said it) or I (implied, I am
   inferring). If implied, quote the line you inferred it from.

3. Contradictions. Places where this person said something that
   conflicts with something else they said in the same call, or
   with a claim I flagged in the input notes.

4. What I did not ask that I should have. One or two questions,
   maximum. Base them on threads the person opened and I did not pull.

Rules. No summary paragraph. No executive summary. No “key
takeaways.” No emojis. If a section is empty, write “None” and stop.

Transcript:
[paste]

My input notes (optional, may be empty):
[paste or leave blank]

Four sections, structured, no summary. The lack of a summary is deliberate. A summary is where the model launders vagueness back into your week. I want the raw material, not a briefing.

4. Cross-call pass, at the end. Once every call for the week has its four-section output, I paste all of them into one document and run a second, much smaller prompt.

Below are per-call synthesis outputs from [N] calls this week.
Roles: [list].

Return only two things.

A. Signals repeated by two or more people, in different calls,
   in different words. For each, cite which calls. Do not list
   signals that only one person raised.

B. Signals raised by exactly one person that I should still take
   seriously, and one sentence on why. Cap at three.

Do not summarize. Do not recommend actions.

That is the last thing a model does for me. What comes out is a short list. Usually five to eight lines total.

The part I still do by hand

The five minutes at the end.

I take the short list from step 4, and I write the one-page synthesis for my PM lead by hand. Not because the model cannot write it. It can. It writes it beautifully. That is the problem.

The one-page synthesis is where I decide what to argue for. Which two signals move a lab priority for next quarter. Which one signal, even if only one person raised it, is the one I am going to spend political capital on. Which contradiction I am going to name in front of the senior director, knowing it will make somebody defensive. That is the work. Not the writing. The load-bearing choices inside the writing.

If I let a model do it, three things happen, and I have tested all three. It flattens the strongest signal into the middle of a balanced list. It softens the contradiction into “an area to explore.” And it removes my name from the decision, because the sentences read like a briefing from nobody. My PM lead is very good at his job. He reads that draft and asks me, politely, what I think. And I have to redo it anyway, slower, because now I have to argue with a version of my own opinion that a model wrote.

So I stopped. The model does the extraction, the structuring, the pattern matching across calls. I do the sentence that says “we should prioritize the lab refresh over the content-developer onboarding, and here is why.” That sentence has a name attached to it. Mine.

The rule I use: if the sentence would be different depending on who wrote it, a human writes it. If the sentence would be the same regardless of author (a quote, a count, a structural note), a model writes it.

Sidebar: three prompt patterns I stopped using

1. “Summarize this call in three bullet points.” The compression is where the signal dies. Three bullets always come back generic, always in the same shape, always missing the one thing that actually moved. I ask for verbatim quotes instead, and I do the compression myself.

2. “What are the key takeaways?” Every model has a house style for takeaways, and it is not mine. “Key takeaways” is also the phrase most likely to trigger the model to invent a fifth bullet because four looks weak. I ask for signals with citations, not takeaways.

3. “Act as a product manager and...” Role prompts sound clever and produce worse output. The model does not become a PM. It becomes a generic idea of a PM, which is exactly the voice you do not want in your synthesis. I tell it what the reader is and what to return, not who to pretend to be.

Why this order

One call at a time before any cross-call pass, because you cannot detect a pattern from calls the model has already blurred together. Per-call output first, then cross-call. The extra 10 minutes at the front saves the whole exercise from producing the kind of synthesis where every week sounds the same.

No summary sections in any prompt, because a summary is a place where the model quietly interpolates. I want the four sections raw, and I want to be the one who decides what they add up to.

Verbatim quotes with attribution, always. When I take a claim into a room on Monday, I want to be able to point at the sentence and the speaker. “The lab partner said this at minute 34” is a very different argument from “customers are frustrated.”

The whole workflow takes me about 40 minutes on a four-call week, most of which is the five minutes at the end, done four times, on the sentences that matter.

What to steal

If you take one thing from this, take the per-call versus cross-call separation. Almost every synthesis prompt I see PMs share online batches everything into one call. It looks efficient. It produces a briefing that reads well and says nothing.

The rest is variations. Different tool, different transcript source, different stakeholder types. The structure holds.

The five minutes at the end holds too. That part is not a workflow problem. It is a job description.

See you next Tuesday.

Luiz

What does your team’s customer-call synthesis look like right now, and where do you still refuse to hand the work to a model? I read every reply, and the best answers usually show up in the comments.

Views are my own and do not represent Microsoft.