name: question-set-builder description: > Read one of your own prospect calls and get your question set handed back to you, marked up. Paste a call transcript — or, if you have no recording, type out the questions you remember asking — and get back: the questions you ask every time, the ones that got a dead answer with what to ask instead, every place you answered for them, your talk ratio in the first 15 minutes when you have a transcript — minutes of their voice when you do not — and the set you should be running next call. One call is enough. You do not need a library. Triggers: "question set builder", "build my question set", "audit my questions", "what should I be asking on my calls", "read my prospect call", or any request to review or rebuild the questions asked on a prospect call. metadata: author: Kathryn Brown, Practice Builders version: "1.1.0" date: "2026-08-11"

Question Set Builder

Your own question set, handed back to you with the dead ones marked and a replacement written. From one call.

Core Principle

Read the whole call before marking anything. A question is not dead because of how it is worded. It is dead because of what came back. That means the answer has to be read before the question can be judged, and the answer three turns later often reveals that a question two minutes earlier never landed. Mark nothing on the first pass. Read the entire input — every turn, start to finish — then go back and score.

Skipping this produces the failure that makes the whole audit worthless: flagging well-worded questions that worked and passing over clumsy questions that opened the person up.

What This Skill Does

You paste one call. Either a transcript, or — if the call was not recorded — the questions you remember asking, typed out from memory. The skill reads the whole thing, scores every question against the 4-Move Question Framework™, and hands back a marked-up version of your own question set.

Job 1: Shows you what you actually ask — Not what you meant to ask. The questions that appear on every call whether you planned them or not, separated from the ones you improvised. Most owners are running a standard question set they have never written down and cannot recite.

Job 2: Names the questions that cost you the call — Every question that got a dead answer, every place you handed over the answer inside the question, and the number on how much of the first 15 minutes was you talking. Each one comes with the replacement question already written.

Job 3: Gives you the set to run next call — A finished, ordered question set built from your own material, with the dead questions removed and the replacements in place. Copy it and run your next call off it.

The Two Input Paths

Both paths are supported. Neither is a fallback. The skill states which path it ran at the top of the output and never apologizes for the path the user chose.

Setup questions. Ask these before producing output, in one message, and wait for the answer.

Transcript path — one question:

1. Which speaker is you?

If the transcript is speaker-labeled and the labels make it obvious, skip the question and state the assumption in the Run Header instead.

Recall path — three questions:

1. Roughly how long is the call?
2. Which of these questions do you ask in the first 15 minutes?
3. For any question where you remember the answer, what did they say?

Never ask more than these. Never ask for a second call. Never ask the user to go find a recording.

The 4-Move Question Framework™

Every question is scored against four moves. This is the rule engine — it is not shown to the user as a lesson, it is applied.

The Framework Boundary — read before scoring anything

Naming the move a question broke is a claim about the question, which is visible in the words the user said and is what the framework is for. Explaining why the other person answered the way they did is a claim about them, which the transcript cannot support. "M2 — the question carried its own answer" stays. "They picked the one they had an answer ready for" goes. When a line could be read either way, rewrite it so the subject is the question.

The Audit: Section by Section

1. Run Header

Orientation. Three lines, no preamble.

Format: Label/value table.

2. Your Question Set

Every question the user asked, deduplicated, grouped into Standing and One-Off.

Standing — asked in a form that will recur. On the transcript path, mark a question Standing when it is structural (opener, scope, budget, next step), when the user signals it is routine, or when it appears more than once in variant form. On the recall path, everything the user typed is Standing by definition — they remembered it because they always ask it.

One-Off — improvised, responsive to something the person said, not part of the pattern.

Say the count out loud. The count is the point: most owners think they ask five questions and are actually running twenty.

The headline is the counts, and it leads: "[N] questions. [M] of them you ask every time." That figure is what surprises people and it stays in the body.

The body shows the Standing set only, capped at 5 to 8 — the ones that open a section of the call rather than follow up inside it. A complete inventory of everything asked is a transcript, which they already have.

Format: Table. Verbatim or near-verbatim wording, not a summary.

Move flags use M1/M2/M3/M4. A dash means clean. Do not explain the flags here — sections 3 and 4 do that.

The full inventory goes to a collapsed appendix at the end — every question, verbatim, numbered, with its flags. Available to anyone who wants it, out of the way of the scan. The word cap below governs the scanned body; the appendix sits outside it.

One call can't prove what's habitual. Say so in a line, and mark the ones that read as routine from where they sit and how they're phrased.

3. Dead Answers

Every question that got a dead answer, with the replacement written.

Counts: a request to repeat or rephrase the question, an answer to a different question than the one asked, a guess the prospect flags as a guess, "I'm drawing a blank," an answer the user has to restate the frame to get, and a shift to an easier topic that never comes back.

Doesn't count: a short answer that is complete, a "no" or "not really" that actually answers, a question the prospect turns around and asks back for a legitimate clarification, or an answer the user simply didn't like. Brevity is not death.

Two people applying this to the same call should land on the same number. When an item is genuinely borderline, leave it out of the count and put it in Section 3 marked as borderline.

On the recall path only: the user says they do not remember what came back. Not remembering the answer is itself a signal.

The count describes what came back. It does not diagnose why. A question can fail to land because it was badly built, and it can fail to land because the line dropped, because the prospect was driving, because they were reading email, or because English is their second language. All of these produce the same transcript. Report the pattern and leave the cause alone — the user was on the call and knows things the transcript does not.

Where a question needed repeating or rephrasing more than once, note it as a landing problem rather than a design problem, and say which. A question that survives a bad connection and a listener working in their second language is a better question either way, so the replacement still stands. The blame does not.

Why it matters: a dead answer is a question that cost you time and returned nothing. Removing it is free. Replacing it is the whole job.

Format: One block per dead question.

The replacement must be a question the user can read off the page on the next call. Not a description of a better question. The actual words.

4. Where You Answered For Them

Every place the user supplied the answer inside the question, or supplied it before the other person could.

This is Move 2, and it is the expensive one. When you hand someone the answer, you have taken them off the hook — they feed it back to you and you have learned nothing.

Detect:

• A question with a candidate answer built in
• A question followed by the asker's own answer in the same turn
• A question the other person answers by agreeing with the asker's framing rather than supplying their own
• A stacked pair where the second question narrows the first into a yes/no

Format: Table.

If there are zero instances, say so in one line and move on. Do not invent one to fill the table.

5. Talk Ratio — First 15 Minutes

A number. Always a number. Both paths produce one.

Transcript path — measured.

1. Find the first 15 minutes. If the transcript is timestamped, use everything up to 15:00 from the first speaking turn. If it is not timestamped, use the first 15/[call length] share of total turns and say so in the method line.
2. Count words spoken by the user in that window. Count words spoken by everyone else.
3. Ratio = user words : other words, normalized to 100.

Recall path — projected from question form.

The shape of a question determines how long the answer runs. Closed and leading questions buy five seconds of their voice. Open questions buy a minute. So the question set itself projects the ratio.

1. Take the questions the user identified as first-15-minute questions.
2. Classify each one:

These three constants are a model, not a measurement. Show them in the output so the user can see and adjust them. They are the skill's assumption about how long each question shape buys, not an observed figure.

1. Their projected answer time = sum of the modeled times. Report it in minutes.
2. Report no percentage on this path. Method line reads: "Modeled from question form. Minutes of their voice, not a share."

Why there is no percentage here. Any share needs a denominator, and every available denominator has to be invented. Share-of-call assumes every second that is not them is the user talking — the window also holds framing, silence and crosstalk, and none of it was measured. Share-of-answer-time needs a constant for how long a question takes to ask, which is a second invented number stacked on the first; at 5, 8 and 12 seconds per question the same set returns 85%, 77% and 70%. A figure that moves 15 points on an assumption nobody set is not a figure.

Minutes need no denominator and lose nothing. "This set buys about 4 minutes of their voice in a 15-minute opening" is a real number, and it moves exactly as hard: swapping 3 closed questions for open ones on an 8-question set takes it from 3.7 to 6.2 minutes — 68% more of their voice, from rewriting three questions. That movement is the whole point of the projection and it survives intact.

Say plainly that this is modeled off the questions, and that it moves the moment the question set changes. That is the point of it.

Measure the window from the first timestamp in the transcript, not from 0:00. Say which numbers you used. Without a counter these are hand counts, so state the method and give a margin rather than implying precision.

Format: Label/value table.

Transcript path — a share. Recall path — minutes, no share.

No target. No gap. No verdict. The opening window is the one that matters — whoever talks most early decides what the call is about. Give the number with no verdict attached, including when it comes back healthy.

Movement is the read, not a benchmark. Where a prior audit exists, show this number beside their own previous one. It moves when the question set changes, and that comparison needs no standard to mean something.

Never skip this section. Never replace the number with "not enough data." If the input is thin, compute from what exists and name the method.

6. The Set You Should Be Running

The rebuilt question set. Ordered, numbered, ready to run. This is the section the user came for.

Build it from:

• The user's own clean questions, kept verbatim
• The replacements written in sections 3 and 4
• Nothing invented that has no root in the input

Order: why they booked → what is happening now → what triggered it → what they have tried → who else decides → what happens if nothing changes → what they want next.

Format: Numbered list. Each item is the question in quotes, followed by a short bracketed note only where the note earns its place.

Cap it. Nine questions maximum. If the audit produced more, cut to the nine that do the most work and list the rest under a short "Hold in reserve" line.

Close the section with the three moves as a running rule, stated once, in the user's terms:

• Ask it open.
• Then stop talking.
• Say back what they said before you move on.

7. The One Change

One line. The single highest-leverage swap out of everything above. Name the question to kill and the question to run in its place.

Format: One sentence. No table, no heading decoration.

Quality Check (Internal — never shown to the user)

Run this silently before presenting. Correct the audit against it. The user never sees a Pass/Fail table.

Enforcement: Run all eight checks. Identify the weakest section, rewrite it, and verify the rewrite is in the final output.

1. Any failed check is fixed before presenting. No flag, no note, no apology in the output.
2. If the input contains fewer than three questions total, say so in the Run Header and run the audit on what exists. Do not refuse.
3. If the replacement questions all sound alike, rewrite the weakest two.

Rules

• One call per run. Not a library, not a quarter's worth. One call is the unit. Running two at once averages away the pattern that makes the audit useful.
• Greetings and logistics are not questions. "How are you?", "Can you hear me?", "Do you want to jump in?" — exclude them from the count and from the set. Count only questions asked to learn something about the buyer or the deal. Say in the Run Header how many were excluded.
• Read everything before scoring. The core principle is a rule, not a suggestion.
• Quote, do not summarize. Questions and answers appear as they were said. A summarized question cannot be audited.
• Every flag needs evidence. A quoted answer, a quoted question, or a counted number. Nothing gets marked on impression.
• Write replacements, not advice. "Ask something more open" is not output. The replacement question, in quotes, ready to read, is output.
• Both paths are first class. Never tell the recall-path user that the output would be better with a recording. Run the audit they asked for.
• Mirror their language. Use the user's own words for their service, their buyer, and their offer. Add no industry vocabulary they did not use.
• No scores, no grades. No letter grade, no "you scored 6/10." The audit shows what happened and what to run instead.
• Never diagnose why an answer fell flat. A transcript shows what was said, not what the other person understood, heard, or was doing at the time.
• Never characterize the prospect. No reads on their sophistication, attention, fluency, or good faith. The audit is about the user's questions.
• Speaker-labeled transcripts split sentences across turns, so the first words of a line often belong to the previous speaker. Verify every quote against the label boundary rather than the visual line.
• Keep the scanned body under 800 words. It gets scanned before a call, not studied. The collapsed appendix sits outside the cap. When the body runs long, cut in this order: standing set entries down to 5, then commentary, then the shortest dead-answer items. Never cut a finding to make room for a section that has nothing in it, and say at the end when findings were held back and how many.
• Always produce a number, always name the method, never dress a modeled number as a measured one. These three coexist; none of them licenses skipping the figure.
• Never name or profile the other person on the call. They are "the prospect" or "they." No demographic read, no assessment of them.
• Keep it scannable. Tables and short lines. Nothing in section 2 through 5 should take more than five seconds to absorb.
• Output as a markdown file. Name it question-set-builder-[YYYY-MM-DD].md (lowercase, hyphens, no spaces).

Output Format

# Question Set Builder

| | |
|---|---|
| **Input path** | [Transcript / Recall] |
| **What was read** | [Description] |
| **You are** | [Speaker label] |
| **Questions found** | [X] |

---

## 1. Your Question Set

**[X] questions. [Y] of them you ask every time.**

| # | Question (as asked) | Standing / One-Off | Move flags |
|---|---------------------|--------------------|------------|
| 1 | "[Question]" | Standing | — |
| 2 | "[Question]" | Standing | M1 |
| 3 | "[Question]" | One-Off | M2, M3 |

---

## 2. Dead Answers

**[X] questions came back empty.**

| | |
|---|---|
| **Question** | "[As asked]" |
| **What came back** | "[Their answer]" |
| **Move broken** | [M1–M4] |
| **Ask instead** | **"[Replacement question]"** |

| | |
|---|---|
| **Question** | "[As asked]" |
| **What came back** | "[Their answer]" |
| **Move broken** | [M1–M4] |
| **Ask instead** | **"[Replacement question]"** |

---

## 3. Where You Answered For Them

**[X] places.**

| # | What you said | What it handed them | Ask instead |
|---|---------------|---------------------|-------------|
| 1 | "[Quote]" | [What they no longer had to produce] | **"[Open replacement]"** |

---

## 4. Talk Ratio — First 15 Minutes

**Transcript path** — a share. **Recall path** — minutes, no share.

| | |
|---|---|
| **Your share** | [X]% — transcript path only |
| **Their share** | [Y]% — transcript path only |
| **Their projected voice** | [N] minutes of a 15-minute opening — recall path only |
| **Method** | [Measured from timestamps, ±[n] pts] — transcript path<br>[Modeled from question form. Minutes of their voice, not a share.] — recall path |
| **Window** | [start]–[end] [offset noted if the recording starts late] |

[Where a prior audit exists: this number beside their own previous one. No target, no gap, no verdict.]

---

## 5. The Set You Should Be Running

1. "[Question]"
2. "[Question]"
3. "[Question]"
4. "[Question]"
5. "[Question]"
6. "[Question]"
7. "[Question]"
8. "[Question]"
9. "[Question]"

**Hold in reserve:** [any cut questions, one line]

Ask it open. Then stop talking. Say back what they said before you move on.

---

## 6. The One Change

[One sentence — the question to kill, the question to run instead.]

What Makes This Different

Ask any AI to "review my sales call" and you get a summary of the call. You already know what happened on the call. You were there.

This reads the call for one thing: your questions, and what came back. It separates the questions you ask every single time from the ones you improvised, which is the first time most owners see their own standard question set written down. It marks the ones that got nothing back — and a question that got nothing back is not a bad question in theory, it is a question that already failed on a real call with a real buyer. Then it writes the replacement, in words you can read off the page tomorrow.

And it counts. Talk ratio in the first 15 minutes is a number, not a feeling. Most owners guess low. The number is where the audit stops being an opinion.

One call is enough. If you have no recording, type out the questions you remember and it works from those — the questions you remember are the ones you always ask, which is exactly the set that needs the audit.

License

Copyright (c) 2026 Kathryn Brown, Practice Builders

This skill is licensed for your personal and business use. You may run this skill inside your own practice and share the outputs it produces with your team and clients. "Your practice" includes employees and contractors engaged to perform work for your business under your direction — virtual assistants, operations support, bookkeepers, and similar team members.

You may not share, distribute, resell, or repackage the skill file itself — including this SKILL.md document, its prompts, frameworks, and structure — with anyone outside your practice. This includes peer practitioners, other consultants who would use it in their own client work, and anyone outside your operating team. Written permission from Kathryn Brown (kathryn@creatingyourplan.com) is required for any redistribution.

This skill is provided "as is" without warranty of any kind, express or implied.